[SQL] Support the MODE aggregate function by desugaring into GROUP BY + COUNT + ARG_MAX - #6806

APPROVE.

Nice desugaring: MODE(v) → inner Aggregate(G∪{v}, COUNT(v)) → outer Aggregate(G, ARG_MAX(v, cnt)), joined back with other aggregates via IS NOT DISTINCT FROM on the group keys. Using COUNT(v) rather than COUNT(*) is what makes the NULL semantics fall out naturally — the all-NULL sub-group gets cnt=0 and only wins when it is the only candidate, so MODE of an all-NULL group is NULL and NULLs never beat a real value. The FILTER case is handled by wrapping the value in CASE(cnt > 0, value, NULL) so groups whose rows are all filtered out survive with a NULL mode. Repeated MODE calls with the same (arg, filterArg) share one branch; the final projection reassembles the original column order. Tests cover the happy path, FILTER (including a group with only NULLs passing the filter and a group with no passing rows), interleaved aggregates with a shared MODE branch, empty input, and DISTINCT rejection.

A couple of small observations, none blocking:

• RejectUnsupportedPlans is a good pattern — worth reusing for other Calcite-accepted-but-unsupported constructs later. Consider giving it its own file once a second case appears.
• The rule silently returns for GROUPING SETS / CUBE / ROLLUP (non-SIMPLE group type). That will fall through to whatever the backend does with a raw MODE, which today is nothing. If you want a clean error rather than a mysterious backend failure, RejectUnsupportedPlans is the natural place to also reject MODE under non-SIMPLE groupings.
• Minor: docs say "The rule for selecting the value is not specified if multiple values are tied for the highest frequency" — worth noting that ties in practice depend on ARG_MAX's ordering (which for MODE is a count, and for equal counts is effectively arbitrary). Not a change request, just a doc-precision thought for future.

LGTM.