vfpu_multi_dot ran the exact dot four times, each summing across the lanes. Now each lane works
out one output, a term at a time, the same steps as vfpu_dot_neon, so nothing crosses lanes until
the rounding at the end. Inf and NaN still take the old way. The IR interpreter (no JIT, as on
iOS) runs every multi-dot through it: GTA 3% fewer cycles, Tekken 2%.
The VFPUDot test checks it against the reference too.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
