Accelerate quantization rescale search with SIMD interval pruning

[gouyt13]

Problem and change

Extra-bit quantization spends substantial time processing rescale candidates through a priority queue. This change adds a float32 AVX2/AVX-512 interval search with conservative pruning and scalar fallback for ambiguous comparisons or work limits. The scalar sweep also replaces pop/push pairs with one heap repair, and successful warmed SIMD searches reuse thread-local scratch buffers.

Validation

Recorded branch validation:

• Portable Release and ASan/UBSan builds: 71 C++ tests passed each, including explicit AVX2 and AVX-512 coverage. LeakSanitizer was disabled under ptrace.
• Rebuilt portable wheel: 102 Python tests passed. Installed CMake consumer passed.
• Formatting, Python lint, whitespace, and configured core/test clang-tidy checks passed.
• Reference checks covered exhaustive objectives, interval bounds, threshold ties, zeros, tails, and threaded scratch reuse. Link-order checks verified backend ISA isolation.
• Strict MkDocs build and git diff --check passed after the documentation update.
  GIST benchmark: 960 dimensions, 96 pinned workers, Intel Xeon Gold 6418H, GCC 14.3.0, portable build; median of five trials, each encoding 24,576 distinct rotated residuals four times. Timing includes code/factor emission and fallback, excluding preparation.
  At eight total bits, throughput increased from 79,280 encodings/s to 184,433 with AVX2 (2.33×) and 372,528 with AVX-512 (4.70×). Across bits 2–9, all 196,608 input/bit cases matched baseline scales, code bytes, and normalization factors.
  These results measure quantization throughput, not full index-build speed. Both backends ran on an AVX-512-capable host; AVX2-only emulation was unavailable.

Compatibility

No changes to public APIs, index formats, code layouts, dependencies, or distance-estimation formulas. The fixed-scale quantization path is unchanged.