Add rerank to primary-key vector search

[JingsongLi]

Summary

Add exact reranking to primary-key vector search so approximate ANN candidates can be reordered
with the original vectors stored in Paimon. The implementation reuses the Data Evolution refine
configuration semantics and supports both local and distributed Spark execution.

Changes

• Share refine/rerank option resolution between Data Evolution and primary-key vector reads,
  including aliases, prefix precedence, query-over-table precedence, factor validation, and
  overflow-safe candidate limits.
• Keep ANN candidates separate from exact fallback candidates, over-fetch only the ANN stream,
  reread selected rows by snapshot-scoped physical position, and recompute exact distances before
  the final global Top-K merge.
• Propagate bounded ANN and exact candidate streams through Spark tasks, then merge and rerank the
  global ANN candidates on the driver without starting a second Spark job.
• Document the shared configuration behavior, recall limitation, and latency/I/O trade-off.

Testing

• Core focused tests: 49 passed, including configuration precedence, aliases, factor 1,
  overflow, candidate-stream isolation, and end-to-end reranking.
• Spark 3 PrimaryKeyVectorSearchTest: 8 passed, including distributed driver-side reranking.
• Compiled changed production modules without the fast-build profile.
• Ran Spotless checks and git diff --check.

Notes

• Reranking is disabled when no refine option is configured.
• A refine factor of 1 recomputes exact scores without over-fetching.
• This change does not modify table, index, or result formats.

[leaves12138]

Reviewed the Core and Spark rerank paths, including candidate merging, snapshot-scoped physical rereads, deletion-vector handling, metric conversion, and refine option compatibility. Focused Core and Spark 3 tests passed locally.