Phase A: Stabilize float32 AMICATorchNG via Kahan + mixed precision #75

Context

Enabler phase of the Apple-Silicon GPU epic. Apple GPUs have no FP64, so all Apple-GPU
acceleration requires a stable float32 AMICA. #70 established that float32 diverges to
NaN on the full 30504-sample data (precision-driven instability, not a divide-by-zero)
and that computing the responsibilities in float64 alone does not fix it.

Approach

• Kahan / compensated summation for the sufficient-stat accumulation (~30504 samples
  per block/component - the prime float32 precision sink Stabilize float32 AMICATorchNG for the GPU fast path #70 never touched). Near-float64
  accuracy with an error bound independent of the term count N, at ~2-3x the sum cost.
• Mixed precision: float64 for the density |y|^rho, the log-likelihood, and the
  stat accumulation; float32 for the matmuls. Guard behind dtype so the default float64
  path is untouched.

Acceptance

• float32 (or mixed-precision) fit converges on the full sample data (32ch x 30504) across

  =5 seeds, Newton on and off, with no NaN.

• Final LL within parity tolerance of the float64 fit.
• Default float64 path stays bit-identical (regression-guarded).
• Tests use real sample data (seed sweep) under tests/torch_tests/.

References

Supersedes #70 (parked). Research: .context/mps_pathways.md Pathway A. Perf-sensitive
regions identified in .context/issue-63/perf_findings.md sections 1 and 4.