[WIP] [lumina] Reuse LuminaSearcher across queries via searcher pool

[chenghuichen]

Purpose

Tests

[JingsongLi]

This PR is marked as [WIP]. A few observations:

1. Searcher Pool design: The LuminaSearcherPool concept is valuable for reusing expensive searcher instances across queries. Key questions:

   • What's the eviction strategy? LRU? Time-based?
   • Is the pool bounded? What happens when it fills up?
   • How does refresh work when the underlying index is updated (e.g., after compaction adds new segments)?

2. Thread safety: Vector index searchers may hold file handles or memory-mapped regions. Ensure the pool handles concurrent access correctly — particularly the case where a searcher is returned to the pool while still being used by another thread.

3. Resource cleanup: When the pool is closed, all pooled searchers must release their native resources (e.g., Lucene IndexReader handles, mmap regions). Is there a close() method on the pool?

4. PR description is empty ("Purpose: / Tests:") — please fill in the design rationale and test plan before requesting review.

Happy to look more closely once this is ready for review.