Fix undefined signed shifts in odd-bit SIMD kernels - #67

Summary

• load packed high-bit planes as uint64_t without alignment or aliasing assumptions
• perform the 3-, 5-, and 7-bit unpacking shifts with unsigned semantics in both AVX2 and AVX-512 kernels
• add a backend-independent regression test that exercises bit 63 and compares both SIMD implementations with a scalar inner product
• remove unused includes and replace two test VLAs with standard fixed-size arrays

Cause

The packed 64-bit high-bit plane was loaded as int64_t. When bit 63 was set, the value became negative and expressions such as top_bit << 1 invoked undefined behavior under C++17, which UBSan reported at runtime.

Verification

• Release C++ tests: 22/22 passed
• ASan + UBSan tests: 22/22 passed
• Python tests: 54/54 passed
• clang-format, clang-tidy, Ruff, ShellCheck, and git diff --check: passed

GIST performance

Pinned single-thread IVF searches over all 1,000 GIST queries, k=10, nprobe=50, seven repetitions:

All result IDs and distances were exactly identical, with maximum distance delta 0.0. The sub-1% QPS variation is consistent with measurement noise. The AVX2 and AVX-512 object text sizes are unchanged, and the helper copies are fully inlined.