fix: release Python GIL and isolate concurrent index searches

[gouyt13]

Summary

Python index calls hold the GIL during native work, while HNSW and SymphonyQG searches update shared search-window state. Release the GIL around native index search, construction, updates, and file I/O, and keep search parameters local to each call so independent Python threads can safely search an unchanged index.

Add per-index access guards: searches, saves, and property reads may coexist; conflicting mutations immediately raise RuntimeError("Index is busy: conflicting operation in progress"). Retain input array owners and reacquire the GIL before Python object handling and guard destruction. Add QuantizedGraph::search_batch_with_ef while preserving existing default-window APIs.

Python signatures/defaults and saved-index formats remain unchanged. Downstream callers must synchronize conflicting updates and must not mutate input arrays during calls. C++ callers still synchronize mutations themselves; dependent C++ binaries should be rebuilt with matching headers/library. Document the concurrency contract and add Python/C++ regression tests. No release is included.

Testing

Reviewed and tested 46160f1 against main (0973bef) on Linux x86-64, GCC 14.3, Python 3.13:

• Python: 968 passed, 2 platform-specific repaired-wheel checks skipped, including the separately run 18 C++ example integration tests.
• Portable C++ (RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF, GoogleTest 1.14.0): 320 passed, 1 conditional skip; independent OpenMPBinding passed. Available generic, AVX2, and AVX-512 HNSW backend checks passed.
• Supplementary concurrency stress: 20 L2/IP/storage/initializer configurations and 600 operations, including 300 save/load/search round trips; outputs match serial references. Access guards recover after native save errors.
• Full core/binding clang-tidy, C++ formatting, Python formatting/lint, strict MkDocs build, and git diff --check: passed.
• Controlled GIST1M QGKMeans A/B (4,096 clusters, 48 threads, 25 iterations, seed 42, exact final assignment; matched Release/native builds; 3 runs per configuration): contiguous-input medians 14.42 s on main and 14.38 s on this branch; strided-input medians 15.26 s and 15.19 s. No training regression observed. All output centroids/assignments match bitwise; WCSS 1,034,787.870122 and recall@10 0.8196 at nprobe=40. This is not a general speedup claim.

Windows, macOS, native ARM64, and ThreadSanitizer were not run locally.