[core][flink][spark] Support PK-vector indexes for postpone buckets

[JingsongLi]

What changed

• Allow primary-key vector indexes in postpone-bucket mode (bucket = -2) while continuing to reject dynamic buckets.
• Capture one snapshot for postpone planning and writer restore.
• Make Flink and Spark background compaction process both bucket -2 rows and real-bucket Level-0 files.
• Build and publish ANN index changes together with compacted data in one strict atomic commit.
• Exclude postponed data from vector search until compaction and document the visibility behavior.

Why

Foreground fixed-bucket writes for postpone tables remain write-only = true. In deletion-vector mode, both real-bucket Level-0 files and bucket -2 files must be compacted before they become visible. The previous postpone compactor only consumed bucket -2, so marker-only Level-0 data could not be compacted and indexed by the background job.

Validation

• 38 focused Core tests, including strict compact snapshot behavior.
• Flink pinned-snapshot source test and mixed postpone/Level-0 PK-vector integration test.
• Spark PrimaryKeyVectorSearchTest: 9 tests.
• Spark PostponeBucketTableTest: 7 tests.
• Spotless and Core/Flink non-fast validation.
• git diff --check.