Smarter dot products

[HEnquist]

Algorithm: combined sinc for multi-channel resampling

The sinc interpolation inner loop has been restructured. Previously, each channel performed N separate SIMD dot products per output frame (one per nearest polyphase point — 4 for Cubic, 3 for Quadratic, 2 for Linear). The new approach builds a single combined sinc filter per frame by linearly blending the nearest polyphase filters with their interpolation weights, then does one dot product per channel against that combined filter.

This trades a one-time per-frame build cost for a cheaper per-channel evaluation. The build step uses SIMD SAXPY (combined += weight * sinc[k]), accelerated with FMA on AVX/NEON and multiply-add on SSE.

The combined path is selected adaptively based on channel count, since for few channels the build overhead outweighs the savings:

• Cubic: combined path from ≥ 2 channels (4 nearest points, break-even at 2)
• Quadratic / Linear: combined path from ≥ 3 channels (fewer nearest points, break-even later)

Performance (vs master, scalar and NEON, measured on Apple Silicon)

Configuration	1 ch	2 ch	4 ch
Cubic f32 (scalar/NEON)	−4% / −1%	−17% / −13%	−40% / −40%
Cubic f64 (scalar/NEON)	+1% / +2%	−3% / −3%	−40% / −35%
Linear f32 (scalar/NEON)	+3% / ~0%	−4% / −3%	−18% / −21%
Linear f64 (scalar/NEON)	+5% / +6%	~0% / +2%	−15% / −11%
Nearest (all)	+1–5%	~0–5%	~0–4%

Nearest mode does not use the combined path and shows small regressions of 1–5%, likely from minor overhead introduced by the architectural refactor (sinc storage changed from packed SIMD vectors to plain Vec<T>).

Other changes

• Sinc storage simplified: polyphase filters are stored as Vec<Vec<T>> instead of packed SIMD register types (Vec<__m256> etc.). Load intrinsics work equally well on unaligned T* pointers, so no functional change, but the code is significantly simpler.
• New correctness tests: 16 parameterised tests verify that the 4-channel combined path produces output matching the 1-channel direct path (within 1e-10) across all four interpolation types, two resample ratios, and both fixed-input and fixed-output modes.