[core] Read primary-key vector results by position

[JingsongLi]

Summary

Adds the physical-row materialization path for primary-key vector search results. Bucket-local ANN and exact hits can now reuse IndexedSplit for distributed transport and be read from their source data files while preserving physical positions and scores.

Changes

• Reuse the existing IndexedSplit contract instead of introducing a vector-specific split: for primary-key raw reads, its rowRanges are physical positions in exactly one data file and its optional scores follow expanded range order.
• Document that IndexedSplit ranges use the coordinate system of their table read path: stable row IDs for Data Evolution and single-file physical positions for primary-key raw reads.
• Add a dedicated PrimaryKeyIndexedSplitRead and provider to validate the single-file contract, expand ranges into physical positions, align optional scores, and route indexed primary-key reads.
• Add PrimaryKeyVectorPositionReader, which filters file batches by selected positions and exposes the physical position and score through ScoreRecordIterator.
• Keep RawFileSplitRead independent of IndexedSplit: it only provides the lower-level single-file read for an already selected position bitmap, while the dedicated indexed reader owns indexed-split semantics.
• Intersect indexed positions with file-index and deletion-vector filtering before reading, so only live matching rows are materialized.
• Keep scalar predicate pushdown, but disable regular TopN and limit pushdown for indexed primary-key results because applying them before position filtering could incorrectly discard ANN hits.

Testing

• Verify existing IndexedSplit serialization remains compatible.
• Verify physical position and score propagation for both single-batch and multi-batch file readers.
• Verify KeyValueTableRead routes IndexedSplit through the dedicated primary-key indexed reader rather than the generic raw provider.
• Verify indexed ranges are converted to the selected-position bitmap passed to the raw file reader.
• Verify non-contiguous physical positions preserve score alignment.
• Verify a regular read limit does not discard ANN hits at higher physical file positions.
• Run the focused Core suite: 9 tests passed.
• Run Checkstyle, Spotless, Maven enforcers, and the focused tests on JDK 8.

Notes

• Follows [core] Add primary-key vector bucket search #8569, which adds bucket-local primary-key vector search and returns data-file plus physical-position results.
• Generating IndexedSplit instances from bucket search, cross-bucket Top-K, snapshot scan routing, and Flink source integration remain follow-up work.

[leaves12138]

Reviewed the latest commit (17ed7d8). The dedicated primary-key indexed read/provider cleanly isolates indexed reads from regular raw-read TopN/limit mapping state, while preserving predicate and deletion-vector filtering. I also verified physical-position/score propagation, multi-batch behavior, serialization compatibility, and the regular-read-then-indexed-read cache order. The focused 22-test suite and the additional cache-order regression passed on JDK 8. Looks good to me.