Add vectorized vec5 cubic and tree32 phase approximations + benchmark updates

[falseywinchnet]

Motivation

• Improve atan2/phase computation performance by adding a vectorized 2‑lane path and a new cubic vec5 first‑octant approximation while keeping the existing tree32 degree7 approach.
• Make the core phase computation generic so different first‑octant strategies can be selected and reused in scalar and vector paths.
• Expand the benchmark to measure multiple phase implementations and sin/cos table performance.

Description

• Added low‑level vector helpers and predicates/macros (V2_DIV, V2_CMPGT, V2_SELECT) and a BRUUN_ALWAYS_INLINE macro to src/detail/bruun_kernel.hpp to support portable 2‑lane SIMD code for SSE2 and NEON.
• Introduced bruun_atan_leaf_cubic, a new bruun_phase_first_octant_vec5_cubic scalar implementation, and retained the previous tree32/degree7 code as bruun_phase_first_octant_tree32_degree7.
• Factored a generic templated bruun_phase_atan2_core that accepts a first‑octant function, added a 2‑lane SIMD version bruun_phase_first_octant_vec5_cubic_v2, and a paired vector entrypoint bruun_phase_atan2_mag_pair that computes two phases at once; replaced bruun_phase_atan2_mag to call through the core path.
• Updated FFT forward_mag_phase to use the paired/vectorized path when available and to write magnitude/phase back into output more directly.
• Updated examples/atan2_benchmark.cpp to include phase in samples and to time multiple variants (std::atan2, tree32, default, vec5 scalar, vec5 pair when BRUUN_LEVEL>=1, float variants, and sin+cos table), and added formatted benchmark output lines.

Testing

• Built the library with -DBRUUN_LEVEL=1 (x86_64 with SSE2) and compiled examples/atan2_benchmark, which executed and produced benchmark output successfully.
• Ran the project's test/validation step (ctest / unit tests) after the build and observed no failures.

---

Codex Task