Enable AVX2 scheduled inverse FFT

[falseywinchnet]

Motivation

• The double-precision d4 inverse fusion keeps q==8 leaves register-resident, reducing leaf store/reload cost and making the flat scheduled inverse more attractive on AVX2.
• Add an inverse-only benchmark column so inverse performance can be measured directly and the AVX2 schedule tradeoff can be evaluated without relying on round-trip inference.

Description

• Enabled the scheduled residue inverse runner for BRUUN_LEVEL 2 (AVX2) by switching the conditional in src/detail/bruun_kernel.hpp so AVX2 uses run_inv_residue_schedule for N>=64 and added a short rationale comment.
• Added a direct inverse timing measurement inverse_ns and printed Inv_ns in the benchmark output by updating examples/benchmark.cpp to time plan.inverse(...) and emit the new column.
• Updated src/detail/inverse_acceleration_notes.txt to record the AVX2 schedule decision, the validation platform/flags, and the rationale that the d4 fusion makes the scheduled inverse beneficial for BRUUN_LEVEL 2.

Testing

• Built and ran automated tests with AVX2 flags via make -j$(nproc) test examples CXXOPTFLAGS='-O3 -mavx2 -mfma -mno-avx512*' and the correctness and api_c tests passed (reported correctness ok backend=avx2-fma-256 and c api ok backend=avx2-fma-256).
• Ran the benchmark (build/examples/benchmark) across sizes and used the new Inv_ns column to compare direct inverse timings, observing an improvement at N=4096 from about ~11071 ns (pre-toggle) to ~10678 ns in the validation sample.
• Generated x86 assembly with make asm-check and produced AVX2 and SSE2 assembly outputs successfully for inspection.
• All automated build/benchmark steps completed successfully in the validation run (no failing tests were observed).

---

Codex Task