Fetch, cache and display electron density and cryo-EM maps - #1134

Exercises the new status checking against the real wwPDB archive: downloadFile
must throw HttpStatusException rather than writing the error page to the
destination, and createValidationFiles must not record a .size for it. Without
the second half the cached error page would pass validateFile, since its
recorded size would match the error body exactly.

It is a general 'has the server got a newer copy' helper with nothing
PDB-specific about it, and other caching code outside this package needs the
same Last-Modified comparison. Protected access only reached subclasses and
the io package itself.

Adds org.biojava.nbio.structure.io.density, which downloads and caches density
maps the same way LocalPDBDirectory caches coordinate files, and hands back a
File that a viewer can contour. Closes biojava#947.

Several services publish density for the PDB and they differ enormously in size
for the same entry, so rather than picking one, sources are tried in order until
one answers. The order is smallest-adequate-first, because the smallest form is
usually perfectly good to look at:

  X-ray:  RCSB density server -> PDBe CCP4 -> PDBe density server
          -> wwPDB map coefficients (disabled by default)
  cryo-EM: RCSB density server -> PDBe density server -> EMDB primary map

For 1cbs a density server slice is about a tenth the size of the equivalent pair
of CCP4 files. For the map behind 6hu9 it is 3.7 MB against a 106 MB primary
map. A size limit, 256 MiB by default, is checked against the size EMDB itself
reports before any of the body is transferred, and exceeding it is not an error:
the chain simply falls back to a smaller representation.

wwPDB map coefficients are supported for completeness, since they are the route
RCSB documents now that edmaps.rcsb.org has shut down, but they are structure
factors rather than a sampled grid and cannot be displayed without a Fourier
transform. They are therefore disabled by default, and DensityFileFormat carries
an isJmolLoadable() flag so that a viewer can refuse them rather than silently
drawing nothing.

Notes on the design:

* A source that has nothing for an entry is skipped and the next is tried, but
  any other transport failure aborts the chain. A network outage must never be
  reported as "this entry has no density". When every source is exhausted,
  NoDensityMapException carries the reason from each one, so a caller can say
  why rather than just that it failed.
* A density server response contains both the 2Fo-Fc and the Fo-Fc blocks, so
  the two kinds share one cache entry instead of downloading the identical file
  twice. Which block to read is a display-time decision.
* Cryo-EM entries are found through their EMDB identifier, looked up from EMDB's
  search API with RCSB as a fallback. That lookup also yields the contour level
  the depositors recommend, which is how an EM map should be contoured; it is
  attached to the result whichever source supplied the voxels. The experimental
  method is never inferred from resolution, which BioJava parses incorrectly for
  some cryo-EM entries (biojava#1000).
* Ccp4Header checks for the MAP stamp at byte 208, so a server that answers with
  an error page and HTTP 200 produces a clean cache miss rather than a corrupt
  cache entry.
* Cached results are fully described by a .meta sidecar, so LOCAL_ONLY requests
  are served without opening a connection.

DemoFetchElectronDensity exercises all three outcomes: an X-ray entry, a cryo-EM
entry resolved through EMDB, and an entry deposited without structure factors.

Adds loadDensityMap and clearDensityMaps to JmolPanel, the single point every
viewer in the module already goes through, so a map fetched by DensityMapCache
can be contoured with one call. Surfaces are given stable ids so they can be
addressed or removed individually, and the state that "Reset Display" restores
is re-saved afterwards; otherwise the button would silently discard the map the
user had just asked for.

Three things had to be established by experiment against Jmol 14.31.10 rather
than assumed, and each changed the implementation:

* Option order in the isosurface command is load-bearing. With "mesh nofill"
  placed before the file name, Jmol accepts the command, reports no error, and
  draws nothing whatsoever. It has to follow the file name.

* A negative sigma does not contour at a negative level. Jmol reserves negative
  sigma for its own internal signalling, so "sigma -3.0" silently contours at
  the default level instead: the intended red lobe of a difference map came out
  identical to the blue one. Difference maps are therefore drawn as a single
  signed surface, which also matches Jmol's own shortcut for them.

* A density server response carries both the 2FO-FC and FO-FC blocks, and Jmol
  chooses between them by testing whether the file NAME contains "&diff=1". That
  marker normally arrives in the URL query string, which a cached local file does
  not have. Appending it to the file URL does not work, and neither does the
  "#diff=1" form the reader's own comment suggests: both were measured and both
  returned the 2FO-FC block or nothing. Embedding the marker in the file name
  does work, so the cache exposes the difference map under a companion name,
  hard-linked to the same bytes where the filesystem allows it.

Verified by contouring the real cached files headlessly: the 2Fo-Fc map at 1
sigma gives cutoff 0.356 over a -1.31 to 3.78 range, and the difference map at 3
sigma gives 0.374 over -0.69 to 0.85 - a different data block, which is what
proves the marker works rather than merely being accepted.

A map that cannot be contoured without a Fourier transform is rejected with an
explanatory exception rather than producing an empty surface.

Puts the feature in reach from the alignment viewer's View menu. The fetch runs
on a SwingWorker: even the smallest source is a few hundred kilobytes and a
full-resolution map can be far larger, so fetching on the event dispatch thread
would freeze the window for the duration.

Requests are built with allowNonRenderableFormats(false), which keeps the map
coefficient source out of the chain automatically rather than relying on the
viewer to notice it cannot draw the result.

When nothing is available the dialog explains why rather than listing HTTP
codes: for an entry whose every source returns 404 the likely reason is that no
structure factors were deposited and there is no associated EMDB map, which is
worth saying plainly. The per-source detail is still shown underneath.

AbstractAlignmentJmol gains getFrame() and setStatus() so a listener in the
neighbouring package can own its dialogs and report progress; both were
previously reachable only as protected fields.

DemoShowElectronDensity displays 1CBS with both maps clipped around the bound
retinoic acid.

Forty-three unit tests run with no network at all. The chain tests use stub
providers, so the behaviour that matters can be pinned exactly: a 404 falls
through to the next source, a too-large map falls through as well, and anything
else aborts. That last one is the point of the design - reporting a dropped
connection as "this entry has no density" would be worse than failing.

The offline set also pins the two-character directory rule against both
spellings of an entry (1cbs and pdb_00001cbs must land in "cb", not "db"), that
every source and kind combination maps to a distinct file, and that a LOCAL_ONLY
request is served entirely from disk with every server pointed at a dead port.

Six integration tests exercise the real services for a couple of megabytes in
total. The cryo-EM path is covered without downloading the 116 MB map: the EMDB
entry is resolved, the author contour level checked against its known value, and
the size guard then declines the full map before any of its body is transferred.
The coefficient test corrupts a downloaded file afterwards to confirm the
ETag-derived MD5 actually catches it rather than merely being recorded.

Writing the header check turned up a real bug: isCcp4 rejected any file shorter
than a CCP4 header, but a gzipped map compresses to a small fraction of the
header it contains, so small EMDB maps would have been rejected as invalid. The
length shortcut is gone; reading decides it.