Phase 4: expose PMI/MIR through the MNE wrapper

[neuromechanist]

Summary

Wires the #133 separation-quality metrics onto the MNE-facing wrapper: AMICAICA.mir(inst, model_idx=...) (Mutual Information Reduction, in nats) and AMICAICA.pmi(inst, model_idx=...) (pairwise mutual information between the fitted sources). Both extract the fitted channels from a Raw/Epochs and delegate to AMICA.mir/pmi (which already exist from #133), so MNE-side users get the same metrics as EEGLAB-side users.

Closes #143. Part of epic #139. This is the final phase.

What's new

• AMICAICA.mir(inst, *, model_idx=0, nbins=None) and pmi(inst, *, model_idx=0, nbins=None). model_idx is keyword-only (consistent with the other consumer methods) and validated against the fitted model count; both guard unfitted/degenerate via _check_fitted. Epochs are concatenated along time.

Test plan

• MNE wrapper (test-mne job, real EEG): mir/pmi reproduce AMICA.mir/pmi on the picked-channel array exactly, per model; pmi shape is (n_components, n_components); both work on Epochs; model_idx bounds and unfitted guards. Round-trip verified: wrapper MIR == array-API MIR (33.54 / 33.58 for the two models).
• Local gates green: ruff, ty check ., mkdocs build --strict, typos, and the mne suite (44) on real sample EEG.

[neuromechanist]

Review (2 Sonnet reviewers: code, tests)

Both confirmed the phase is thin, correct wiring; the code reviewer found nothing above its confidence bar (traced the full AMICAICA -> AMICA -> AMICATorchNG delegation chain, ran all 44 tests, verified the round trip). Test reviewer confirmed the per-model comparison genuinely discriminates (model 0 MIR 33.46 vs model 1 33.55). Actionable findings addressed:

• pmi on Epochs was only shape-checked, not value-pinned, so a bug in _data_for's Epochs branch would slip through for pmi (caught for mir). Now value-pinned against AMICA.pmi.
• Added an nbins passthrough test (matches AMICA with a non-default nbins, so a dropped forward is caught) and a degenerate-fit refusal test for mir/pmi (matching the convention used for the other consumer groups).
• Made per-model discrimination explicit (assert not allclose(model0, model1)), self-documenting the model_idx-honored invariant.
• Added mir's missing return annotation.

Local gates green: ruff, ty check ., mkdocs build --strict, typos, and the mne suite (46) on real sample EEG.