AVX2: eight-wide f32 BODFT forward/inverse combines, FMA intrinsics and SIMD tail scheduling improvements

[falseywinchnet]

Motivation

• Improve BODFT performance by adding an AVX2 eight-position single-precision forward and inverse combine so f32 work uses 256-bit lanes instead of falling back to 128-bit paths.
• Reduce inverse overhead and tail penalties by tightening SIMD tail flow and running 128-bit V2/V4F tails after wider AVX2/AVX-512 loops.
• Reduce instruction count and live pressure by preferring accumulator-first FMA intrinsics where available and folding constant 0.5 factors into twiddle constants for rotations.
• Document kernel dependency/frontier decisions to guide further register-pressure-sensitive optimizations.

Description

• Added AVX2 f32 helpers and eight-wide combine routines: bodft_uzp8_ps, bodft_store8_ps, bodft_rev8_ps, bodft_combine_fwd_avx2_f32, and bodft_combine_inv_avx2_f32, and wired them into combine_fwd/combine_inv dispatch for float (sizeof(RT)==4).
• Updated x86 SIMD macros in bruun_kernel.hpp to use _mm_fmadd_* / _mm_fnmadd_* when __FMA__ is defined (falling back to separate mul+add/sub otherwise), and updated 128-bit float/double codelets to shorten live ranges and move loads after rotation start where profitable.
• Reordered several AVX2/128-bit loops so wide SIMD runs first and 128-bit tails run afterwards, and folded the 0.5 scale into twiddle constants in float inverse loops to remove redundant half multiplies.
• Added documentation and notes: new kernel_sequentiality_flow.md, updates to inverse_acceleration_notes.txt, and BODFT AVX2 notes (bodft_avx2_notes.md) describing validation and observed performance changes.

Testing

• Ran the full test suite with AVX2/FMA: make test BUILD_DIR=build-avx2 CXXWARNFLAGS='-Wall -Wextra -Wpedantic -Werror', and the CTest checks passed.
• Ran the standalone BODFT benchmark: make build-avx2/examples/bodft_benchmark BUILD_DIR=build-avx2 && build-avx2/examples/bodft_benchmark, and observed measured improvements (example: double inverse ~748 us -> ~611 us at N=65536; f32 inverse ~1.13 ms -> ~0.78 ms at N=65536).
• Verified assembly with make asm-check BUILD_DIR=build-avx2 and inspected build-avx2/src/bodft_avx2.s for AVX/FMA instructions.
• Also built/checked an SSE2/no-AVX configuration to confirm the 128-bit fallback paths remain correct.

---

Codex Task