Replace FHT assembly with AVX intrinsics

[gouyt13]

Summary

Replace the GNU inline assembly and Windows MASM FHT kernels with shared AVX intrinsics used by the AVX2 and AVX-512 rotation backends. Preserve butterfly ordering, four rotation passes, normalization, and saved rotation state while removing the assembler build dependency.

• Support input dimensions from 64 through 65,536, including padded dimensions, and update HNSW load validation accordingly.
• Add scalar-reference/backend parity, norm/inner-product, and large-dimension persistence coverage.
• Remove the obsolete assembly header, MASM snapshot, and generator; update documentation and lint exclusions.

Saved-index formats and production dependencies are unchanged. The previously installed internal rabitqlib/utils/fht_avx.hpp header is removed; downstream code directly including it must migrate.

Testing

• All 173 C++ tests passed with RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF, including explicit AVX2 and AVX-512 rotation checks.
• Installed CMake package consumer built and passed after assembly removal.
• C++ formatting, Python formatting/lint, ShellCheck on changed scripts, focused clang-tidy for both rotation backends, and git diff --check passed.
• Windows/MSVC, the full pre-merge clang-tidy run, and strict MkDocs validation were not run locally.

Local FHT microbenchmarks were kept outside the repository. This PR makes no end-to-end query performance claim.