feat(index): Support vector index rerank

[JingsongLi]

Purpose

This pull request adds rerank support for vector index search, aligned with the Java implementation. When a positive refine factor is configured, the index search first retrieves a larger candidate set and then reranks those candidates by reading raw vectors and computing exact scores.

Brief change log

• Support vector index rerank with refine_factor / refine-factor / rerank_factor / rerank-factor option aliases.
• Resolve rerank options from search options first, then table options, with field/index-specific prefixes taking precedence over global options.
• Expand index search limit to limit * refine_factor and rerank candidates back to the original query limit.
• Read raw vectors for the union of candidate row IDs while scoring only the candidates relevant to each query.
• Keep Java-compatible ranking semantics: higher score first, and smaller row id first when scores tie.
• Optimize rerank performance with bounded top-K heaps, row-id-to-query routing, zero-copy Arrow vector scoring, and streaming per-query top-K collection.

Tests

• cargo fmt --all
• cargo test -p paimon vector_search --lib
• cargo check -p paimon --all-targets
• git diff --check

API and Format

This change adds vector search option aliases for rerank/refine factor. It does not change the storage format.

Documentation

No user-facing documentation is updated in this pull request. The supported options are covered by unit tests.

[leaves12138]

LGTM. The rerank implementation is aligned with the Java behavior, keeps the expected score ordering semantics, and the follow-up optimization avoids unnecessary allocations and full sorts. The accidental README change has been removed, and the check job is now passing.