Add vec5+cubic vectoring phase path and benchmark comparisons

[falseywinchnet]

Motivation

• Introduce a faster vec5+cubic vectoring-first-octant phase approximation as an alternative to the existing tree32/degree7 path for polar phase extraction in the mag/phase kernel.
• Keep the original tree32 path available so both codepaths can be benchmarked and compared in-kernel.

Description

• Added bruun_atan_leaf_cubic residual helper and preserved the original tree32/degree7 function as bruun_phase_first_octant_tree32_degree7.
• Implemented bruun_phase_vec_step and bruun_phase_first_octant_vec5_cubic implementing the vec5+cubic vectoring network, and routed the default bruun_phase_first_octant to the vec5 path.
• Added a generic bruun_phase_atan2_core template to call a chosen first-octant implementation so the tree32 and vec5 paths can both be exercised without removing either.
• Expanded examples/atan2_benchmark.cpp to time and report std::atan2, the old tree32 path, the new vec5 path, and sin/cos vs table sin/cos timings.

Testing

• Configured the tree and benchmark build with cmake -S . -B build -DBFFT_BUILD_SHARED=OFF -DBFFT_BUILD_PROBES=OFF which completed successfully.
• Built targets with cmake --build build --target correctness atan2_benchmark -j2 and the build completed successfully.
• Ran unit/check correctness with ./build/correctness which printed correctness ok backend=avx2-fma-256 (success).
• Ran the benchmark ./build/atan2_benchmark 1024 which executed and reported timings for std::atan2, bfft tree32, bfft vec5, and sin/cos variants (successful run).

---

Codex Task