[core] Search primary-key vector indexes

[JingsongLi]

Summary

Adds the complete Core path for single-query vector search on primary-key tables. A query captures one snapshot, searches bucket-local ANN payloads with exact fallback for uncovered compact files, merges a deterministic global Top-K, and materializes the selected physical rows through the existing IndexedSplit read path.

Changes

• Plan one snapshot consistently across data files, deletion vectors, and active vector payloads, grouped by (partition, bucket).
• Execute PrimaryKeyVectorBucketSearch for every planned bucket, applying snapshot deletion vectors to both ANN and exact fallback results.
• Merge bucket candidates into a deterministic global Top-K by distance and physical location.
• Convert (partition, bucket, file, position, distance) candidates into one-file IndexedSplits with compressed physical ranges and scores aligned to expanded range order.
• Let DataTableBatchScan consume the physical result directly at the captured snapshot instead of scanning a newer snapshot.
• Introduce GlobalIndexSplitResult so DataTableBatchScan depends only on the generic snapshotId() and splits() capability, not on the primary-key vector implementation.
• Route VectorSearchBuilder automatically to the primary-key vector path when the selected field has a configured primary-key vector index.
• Preserve compatibility with delegated table wrappers when constructing the key-value vector reader.
• Rename the existing data-evolution implementations to DataEvolutionVectorScan and DataEvolutionVectorRead, including Spark inheritance and batch scan routing.

API and behavior

The public API remains unchanged:

GlobalIndexResult result =
        table.newVectorSearchBuilder()
                .withVectorColumn("embedding")
                .withVector(query)
                .withLimit(k)
                .executeLocal();

TableScan.Plan readPlan =
        table.newReadBuilder().newScan().withGlobalIndexResult(result).plan();

For a configured primary-key vector field, executeLocal() returns a snapshot-scoped physical result. Passing that result to the table scan produces existing IndexedSplits; no new table split or row-id namespace is introduced.

Partition filters are supported. Non-partition scalar filters are rejected for this path because they cannot currently be applied before bucket Top-K without changing search correctness. Batch vector search remains outside this PR.

Testing

• Merge fragmented data splits into complete bucket inputs while preserving deletion-vector alignment.
• Keep data files and vector payloads on the same captured snapshot.
• Filter payloads by vector field id and index type.
• Preserve bucket search inputs through Java serialization.
• Merge global Top-K deterministically.
• Compress physical positions and align scores and deletion files in IndexedSplit.
• Run an end-to-end public builder query through synchronous index creation, ANN search, DataTableBatchScan, and physical row reading.
• Run 14 focused Core tests covering scan, bucket search, result conversion, indexed split reads, position reads, raw-file reads, and end-to-end search.
• Run all 33 VectorSearchBuilderTest cases plus the primary-key end-to-end test after the rename.
• Run Checkstyle, Spotless, Maven enforcers, compilation, and focused tests without -Pfast-build.

Notes

• Follows [core] Read primary-key vector results by position #8576, which added primary-key physical-position reading for IndexedSplit.
• This PR has 1,421 additions and 20 deletions across 16 changed files and stays within the agreed approximately 2,000-line review scope.

[leaves12138]

Reviewed the latest Core head (80e1f79). The snapshot-scoped planning, deletion-vector alignment, bucket-local ANN/exact fallback, deterministic global Top-K, and physical-position materialization are consistent. The new GlobalIndexSplitResult abstraction also cleanly removes the batch-scan dependency on the concrete primary-key vector result. I ran the focused 11-test Core suite on JDK 8 with Checkstyle, Spotless, Maven enforcers, compilation, and packaging; all passed. Looks good to me.