Add AVX2 forward-combine path to BODFT kernel

[falseywinchnet]

Motivation

• Improve BODFT (odd-frequency DFT) forward performance to match the FFT codepath by adding a 256-bit AVX2/FMA double-precision combine implementation and enable assembly inspection and notes tracking.
• Reduce the twiddle-multiply cost and increase vector utilization by handling four paired radix-4 positions per iteration in the AVX2 path.

Description

• Implemented a 256-bit AVX2/FMA forward combine: added bodft_combine_fwd_avx2_f64 and helpers bodft_uzp4_pd, bodft_store4_pd, bodft_rev4_pd in src/detail/bodft_kernel.hpp and wired it ahead of the existing 128-bit SIMD fallback in combine_fwd.
• Kept the existing 128-bit SoA SIMD paths (SSE2/NEON) and scalar tail for non-AVX2/tail cases, and left the float path on the portable 128-bit flow pending an 8-position AVX2 float pack pass.
• Extended Makefile to emit a BODFT AVX2 assembly target (build/.../src/bodft_avx2.s) and include it in asm-check.
• Added src/detail/bodft_avx2_notes.md to record validation commands, assembly inspection steps, performance observations and next work items.

Testing

• Ran make test BUILD_DIR=build-avx2 CXXWARNFLAGS='-Wall -Wextra -Wpedantic -Werror' and saw both correctness checks pass (correctness ok and c api ok).
• Built and ran the BODFT-specific benchmark with make build-avx2/examples/bodft_benchmark BUILD_DIR=build-avx2 && build-avx2/examples/bodft_benchmark, which produced correctness output and timing numbers for f64/f32 (completed successfully).
• Generated and inspected AVX2 assembly with make asm-check BUILD_DIR=build-avx2 and verified AVX/FMA/ymm patterns via rg 'vfmadd|vfmsub|vperm|ymm' build-avx2/src/bodft_avx2.s, confirming vectorized code emission.
• Built and ran a BFFT benchmark example with make build-avx2/examples/benchmark BUILD_DIR=build-avx2 && build-avx2/examples/benchmark 65536 200 to compare end-to-end timings in this environment (completed; platform-specific comparison columns such as FFTW/MKL were unavailable).

---

Codex Task