Fix HNSW entry-point races and construction failure handling

[gouyt13]

Summary

Fix four correctness failures in HNSW construction and insertion: a parallel entry-point race, invalid cleanup after an initial build failure, corruption after rejecting a second build, and undefined behavior for M = 0 or M = 1.

• Read the entry point and maximum level together under the existing mutex in both raw and quantized insertion. This adds no lock acquisitions.
• Initialize each reserved link pointer before an allocation can fail, own checked-out construction visited sets with RAII, and discard partial construction state after workers join. A failed initial build leaves an empty index that can be built again.
• Reject construction on a built or loaded index before modifying it. Publish Python cluster metadata only after successful construction, preserving the original index after a rejected rebuild.
• Reject M < 2 during construction and loading. Add C++ allocation-failure/retry tests and C++/Python regressions for invalid degrees and rejected rebuilds.

Public signatures, serialized layouts, and dependencies are unchanged. Previously accepted index files with M = 1 are now rejected. Scheduling and numerical formulas are unchanged; Python cluster-ID conversion is outside this change.

Testing

• Full C++ CTest suite: 295 passed, with ASan/UBSan and RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF.
• Python HNSW, quantization factors, portable indexes, thread counts, and persistence paths: 241 passed against a rebuilt and installed wheel; imported extension path and wheel contents verified.
• Both targeted ThreadSanitizer reproducers passed: parallel construction and parallel add(), each with 800 points, eight workers, and forced delays after unlocking the entry-point mutex. HNSW header paths were instrumented; the linked compiled core was not TSan-instrumented. Deadlock detection was disabled for these race-focused checks.
• Original allocation-failure/destruction and rejected-rebuild reproducers passed under ASan/UBSan.
• Full clang-tidy over all 37 configured production translation units, C++ formatting, Python formatting/lint, and git diff --check passed.

Validation ran on Linux x86-64 with GCC 14.3.0. LeakSanitizer was disabled because of the tracing sandbox. ARM64 was not executed. This is a correctness fix; no benchmark or performance improvement is claimed.