[core] Avoid global index rebuilds for unrelated column compaction

[leaves12138]

What

• Make data-evolution global-index refresh planning compare the indexed columns' maximum sequence numbers instead of the compacted file's overall maximum sequence.
• Apply the same decision to both materialized manifest entries and the projected binary manifest scan.
• Fall back conservatively to the file-level maximum for legacy or malformed column sequence metadata.
• Cache table schemas and physical file-field layouts while planning refreshes.
• Project the persisted _WRITE_COLS_SEQUENCES field in the binary manifest path.

Why

Data-evolution compaction rewrites a full normal data file. The file-level maxSequenceNumber advances even when only a non-index column changed, so the previous planner rebuilt an otherwise valid global index.

PR #9313 persists per-column sequence watermarks in DataFileMeta. This PR consumes those watermarks so non-index-column compaction does not trigger index work, while indexed-column updates still refresh the exact affected row range.

The decision no longer depends on retained snapshot history, so it remains correct after snapshots expire and across repeated compactions.

Compatibility

• Legacy files without column sequence metadata remain conservative and refresh based on the file-level maximum sequence.
• Missing or malformed positional arrays also fall back to the file-level maximum.
• Indexed fields absent from a partial-write file do not mark the corresponding index as changed.
• The existing CompactFileIdentifierSet path remains unchanged for recognizing deleted files during projected manifest scans.

Scope

This is the functional follow-up to merged PR #9313.

It does not change DataFileMeta, compaction metadata generation, serializers, split protocols, or CollectedDeletes.

Tests

• DataEvolutionGlobalIndexRefreshPlannerTest covers compacted full files, writeCols ordering, row-tracking fields, malformed metadata fallback, and schema caching.
• SortedGlobalIndexScannerTest covers sequential non-index/index-column updates with compaction, exact refresh row ranges, and snapshot expiration.
• Materialized-entry and projected binary-manifest planning remain in parity.
• GenericGlobalIndexScannerTest.
• Core Spotless and Checkstyle.

[leaves12138]

I have one question:
Does "SplitSerializer" need to add its version by one?
SplitSerializer is used for converting data between java and rust, so in this way, it should be version2-. And SplitSerializer is used for serializing state for some split like FallbackSplit, so it need to be compatible with version-1.
Please review code for SplitSerializer, thanks~

[JingsongLi]

Converted to draft, "Add an optional positional long[] of per-column maximum sequence numbers to DataFileMeta" should be a separate PR.

[JingsongLi]

+1