[server] Support update_if_changed merge engine for primary-key tables

[litiliu]

Summary

Introduce an UPDATE_IF_CHANGED merge engine for primary-key tables ('table.merge-engine' = 'update_if_changed').

It keeps last-row upsert semantics but suppresses value-identical writes:

Stored row	Incoming operation	Result
Absent	Insert/upsert	Store the row, emit an insert changelog record
Present	Every logical field is equal	Keep the stored row, emit no changelog record (no-op)
Present	At least one field differs	Store the row, emit a normal update changelog record
Present	Delete	Delete the row, emit a delete changelog record
Absent	Delete	No-op

Equality is based on logical field values, not raw bytes: null values use logical equality, binary values are compared by content, and rows from different schema versions are aligned to the latest schema by stable column IDs. Fields that exist in the latest schema but are absent from either compared row are treated as null. Partial updates and partial deletes are first applied to the stored row to produce a complete candidate row, which is then compared with the stored row.

Because it must read the stored value before deciding, it does not use the WAL full-row fast path that skips old-value lookup (the merger is not a DefaultRowMerger, so canSkipOldValueLookup stays false).

Fixes #4343

Changes

• MergeEngineType.UPDATE_IF_CHANGED + fromString and config description
• New UpdateIfChangedRowMerger, supporting full-row, partial-update, and partial-delete mutations with logical row equality, wired into RowMerger.create()
• Schema-version alignment by stable column IDs, including schemas with dropped columns
• FlinkTableSink: allow partial updates and UPDATE/DELETE for this engine
• Docs: new merge-engine page + listings/options updates

Test Plan

• UpdateIfChangedRowMergerTest and RowMergerCreateTest (unit: logical equality, schema alignment, full-row and partial mutations, and merger creation)
• FlussTableITCase#testUpdateIfChangedMergeEngine (end-to-end: value-identical full-row upserts emit no changelog record while changes and deletes do)
• FlussTableITCase#testUpdateIfChangedMergeEngineWithPartialUpdate (end-to-end: value-identical partial updates/deletes are no-ops while changed mutations emit changelog records)
• Both client end-to-end tests run with FULL and WAL changelog images
• FlinkTableSinkITCase#testUpdateIfChangedMergeEngineSupportsMutations (Flink sink integration: partial INSERT, UPDATE, and DELETE)

🤖 AI-assisted changes - reviewed by human developer

[litiliu]

@luoyuxia PTAL, when you have time thanks!