Add AVX2 float BODFT forward combine (8-position, FMA)

[falseywinchnet]

Motivation

• Close the remaining performance gap by adding an AVX2/FMA single-precision forward combine so float BODFT can use a 256-bit path instead of only the 128-bit SSE2/NEON path.
• Keep the scalar and 128-bit SIMD paths as fallbacks so correctness and non-AVX builds are preserved.

Description

• Implemented an eight-position AVX2 single-precision forward combine in src/detail/bodft_kernel.hpp with helpers bodft_uzp8_ps, bodft_store8_ps, bodft_rev8_ps and bodft_combine_fwd_avx2_f32 that use AVX2/FMA intrinsics to deinterleave, compute three child twiddle products with vfmaddps/vfmsubps, permute lane-reversed partner bins, and store outputs through an AVX2 packing helper.
• Routed float combine dispatch so when BRUUN_LEVEL >= 2 and sizeof(RT) == 4 the new AVX2 float path is used, while the existing 128-bit SSE2/NEON float combine remains the fallback and scalar tail loop remains unchanged.
• Tightened the 128-bit fallback call to only run when the AVX2 path did not handle any positions to avoid redundant work.
• Documented the change and validation in src/detail/bodft_avx2_notes.md describing the new float combine, FMA usage, partner-bin permutes, fallback behavior, and validation steps.

Testing

• Built the project with cmake --build build -j2 and the build completed successfully.
• Ran unit/functional tests with ctest --test-dir build --output-on-failure and all tests passed (2/2 tests passed).
• Emitted assembly with make asm-check BUILD_DIR=build and verified AVX2/FMA instructions and permutes were present by searching for vfmadd*ps, vfmsub*ps, vpermd/permutes in the generated build/src/bodft_avx2.s.
• Built and executed the standalone BODFT benchmark (examples/bodft_benchmark) on an AVX2/FMA build and on an SSE2/no-AVX build to validate correctness and timings; the benchmark completed and printed correctness/timing output for both float and double runs.

---

Codex Task