Phase D: Multi-model support for AMICAMLXNG (Apple GPU) #81

Context

Phase B (#77) measured that the MLX backend is a ~7x Apple-GPU win for single-model
AMICA (and beats an RTX 4090 at EEG scale), while multi-model has no GPU path: AMICAMLXNG
(v1 MVP, #76) is single-model only, and PyTorch-MPS loses to CPU. Extending the MLX backend to
multi-model carries the ~7x win to real multi-model AMICA -- the highest-value follow-up, and the
last piece before epic #74 closes.

Scope

Add n_models > 1 support to AMICAMLXNG, porting the multi-model machinery from the PyTorch
backend (AMICATorchNG, torch_impl/core.py):

• comp_list indirection (shapes go from (n_mix, n_channels) to (n_mix, n_comps)), per-model
  W = inv(A[:, comp_list[:, h]]) and slogdet on the CPU stream (hoisted per iteration).
• E-step per-model loop producing logV (batch, n_models), cross-model responsibilities
  v = softmax(logV, over models), and u = v*z.
• M-step: gm = sum_t v_h / n, the per-model exact-EM bias c update (Fortran update_c),
  and the gm-weighted A-update scattered through comp_list (byte-identical to single-model when
  n_models=1).

Component sharing (share_comps) stays deferred (a further follow-up; it builds on this).

Acceptance

• AMICAMLXNG(n_models=2) fits the real sample EEG on the Apple GPU (float32) and its converged
  LL / per-block sufficient stats match AMICATorchNG(dtype=float32, n_models=2) within the
  established multi-model tolerance (multi-model is not partition-identifiable, so compare the
  same-init one-iteration stats tightly and the converged LL within distributional spread).
• Single-model (Phase C: MLX backend port for AMICATorchNG (Apple-native GPU) #76) stays byte-for-byte unchanged (guard the n_models=1 path).
• The Phase B benchmark's multi-model configs then run MLX too (drop the single-model gate).

Dependencies

Builds on #76 (single-model MLX). Part of epic #74 (Phase D).