[core] Add bucket-level primary-key vector index maintenance

[JingsongLi]

Summary

Builds the bucket-local maintenance layer for primary-key vector indexes on top of #8549. The maintainer restores active ANN state, reads complete vectors from compact data files, and produces index-file increments for compaction transitions without activating the write or query paths yet.

Changes

• Restore immutable bucket index state from active ANN payloads and validate vector field, index type, payload uniqueness, and one-to-one source-file coverage.
• Read vector-only projections from physical compact data-file rows without deletion-vector filtering, preserving ordinals for later search-time filtering.
• Maintain ANN payloads across compaction by removing segments that reference replaced files and rebuilding one multi-source segment for uncovered active sources.
• Restrict eligible sources to COMPACT files above level 0, where rows are complete for deduplicate and partial-update tables.
• Allow source-backed ANN segments to coexist in index manifests even though their ordinal ranges are segment-local and overlap.
• Clean up newly created ANN payloads when a maintenance attempt fails.

Testing

• Build a single-source ANN segment from a compact data file.
• Rebuild a multi-source ANN segment after partial compaction.
• Reject duplicate source coverage and mismatched vector index types during restore.
• Read vectors from a real Avro compact data file with vector-only projection while preserving NULL physical positions.
• Combine source-backed manifest entries with overlapping segment-local ordinal ranges.
• Preserve ordinary global-index overlap validation in mixed source-backed/range-backed manifests.
• Run the focused Core suite: 15 tests passed.
• Run the non-fast Core reactor build with Checkstyle, Spotless, and Maven enforcers.

Notes

• Follows [core] Add primary-key vector index foundation #8549, which introduced the primary-key vector index metadata and configuration foundation.
• This PR intentionally does not wire the maintainer into the write lifecycle and does not add query routing. Those are follow-up changes.

[leaves12138]

I found one manifest correctness regression. The PR's focused 14-test suite passes, but a mixed source-backed/range-backed regression test reproduces the issue.

[leaves12138]

Re-reviewed the latest head. The source-backed exemption is now scoped to source-backed pairs, the mixed-entry regression is covered, and the focused 15-test suite passes locally. LGTM.