feat: Fix FlatMapFieldWriter::ingestRow empty-map round-trip (T271900360) (#754)

[apurva-meta]

Summary:

Stacks on top of D105703824 to land the primary fix for T271900360.

Problem:
FlatMapFieldWriter::ingestRow re-encodes a ROW vector (the struct
projection a flatmap-as-struct reader produces) back into a flatmap.
The pre-fix path unconditionally invoked the no-inMap overload of
FlatMapPassthroughValueFieldWriter::write, which sets inMap=1 for
every key at every non-null parent row. Result: a (non-null empty map)
row becomes (non-null map of N NULL-valued entries) on disk; the
round-trip is lossy and Vader's checksum validation flags it.

Fix:
For each feature column, compute a per-row inMap bitmap from the
feature's null pattern (NULL value → inMap=0, non-null → inMap=1) and
route through the existing 3-arg with-inMap overload. Empty/missing
keys are now correctly encoded as 'key absent at this row'.

Stats accounting lockstep:
The writer's consistency check enforces
fileRawSize == columnStats.front().getLogicalSize(). To keep that
invariant intact under the new per-feature-presence accounting:

• FieldWriter.cpp ingestRow pre-computes totalKeyCount as the
  sum of per-feature non-null entries within childRanges and
  passes it to collectKeyStatistics / collectPassthroughStringKeyStatistics.
• collectPassthroughStringKeyStatistics takes the per-feature
  values + childRanges and computes the per-key string-length
  contribution scoped to present entries only.
• RawSizeUtils.cpp getRawSizeForPassthroughFlatMap counts only
  present (non-null per-feature) entries for both keys and values.

Test changes:

• dwio/tools/test/TestFormatConverterFlatmapAsStruct.cpp adds
  EmptyMapRowsPreservedThroughIngestRow, a self-contained repro
  that builds a ROW projection of an empty-map row, feeds it
  through the flatmap writer, and asserts the as-MAP checksum
  matches a reference written from a MapVector. Fails pre-fix,
  passes post-fix.
• dwio/nimble/velox/tests/FieldWriterStatsTest.cpp: the three
  pre-existing flatMapPassThrough*FieldWriterStats tests encoded
  the old 'all keys present at every non-null flatmap row'
  accounting via magic numerical stats. Their data setup is
  preserved, but their assertions are restructured to check
  semantic invariants that the new accounting guarantees:
  (1) valueStat.nullCount == 0
  (NULL per-feature values become inMap=0, not value=NULL)
  (2) keyStat.valueCount == valueStat.valueCount
  (each present key has exactly one written value)
  (3) keyStat.valueCount <= numKeys * numNonNullFlatmapRows
  (upper bound from pre-fix accounting)
  (4) keyStat.valueCount > 0
  These invariants are sufficient to detect regression while not
  encoding encoding-dependent byte/physical-size formulas.

Differential Revision: D105712499

[meta-codesync]

@apurva-meta has exported this pull request. If you are a Meta employee, you can view the originating Diff in D105712499.