Close per-query CAGRA matrices in cuvs-lucene

[nvzm123]

Closes #2621.

Summary

• closes each successfully built device-backed CAGRA query matrix immediately after synchronous search result materialization;
• preserves the existing CAGRA/brute-force routing, filtering, scoring, and public APIs;
• adds a real-GPU regression that runs 128 repeated CAGRA queries, verifies the GPU reader and loaded CAGRA index, checks rank-one self matches and unique hits, and confirms workspace allocations are balanced.

Validation

• mvn -Dtest=TestCuVSQueryMatrixLifecycle spotless:check test — 1 test passed
• mvn spotless:check verify — 326 tests passed, 30 skipped

The full run retains existing native-access, Vector API, graph-parameter clamp, dataset-dimension, and GDS/KvikIO fallback warnings.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.