Implement save/load persistence for AMICATorchNG

[neuromechanist]

Closes #36. Also closes the save/load half of #15 (test coverage).

Problem

AMICA.save() / AMICA.load() raised NotImplementedError after #32 removed the old AMICATorch state_dict path. There was no way to persist a fitted PyTorch model.

Change

• AMICATorchNG.state_dict() serializes the model as three parts: config (constructor args, dtype stored by name), params (fitted tensors moved to CPU: A/W/c/mu/alpha/beta/rho/gm/comp_list/mean/sphere), and extra (sldet, iteration, ll_history, optimizer state). Every value is a tensor or plain Python primitive, so the payload round-trips through torch.save/torch.load with weights_only=True (no custom classes or torch.dtype objects).
• AMICATorchNG.from_state_dict(state, device=None) rebuilds the object via the constructor and restores the tensors/scalars. Raises on an unfitted state_dict() and on an unknown format_version.
• AMICA.save() writes the wrapper config + backend state; AMICA.load() is now a classmethod returning a fitted, transform-ready model. The MPS/float64 device fallback used by fit() was factored into _select_device() and reused by load().

Tests (real sample EEG, no mocks)

Replaces the two NotImplementedError placeholder tests with:

• test_ng_save_load_roundtrip - fit -> save -> load reproduces mixing/unmixing matrices and transform() output exactly (lossless CPU float64 restore).
• test_ng_save_requires_fit - saving an unfitted model raises.
• test_ng_load_rejects_unknown_version - a tampered format_version fails loudly (no silent-failure).

Full torch suite: 26 passed.

Notes

• No backward-compat shim for the old (never-working) format.
• load() changed from an instance method to a classmethod (the old body only raised, so no callers relied on the instance form).

[neuromechanist]

Review pass (4 Sonnet reviewers) — all findings addressed

Ran the pr-review-toolkit reviewers (code, silent-failure, test-coverage, comments) on the diff. Every material finding was addressed in commit 8746236; no false positives skipped.

Silent-failure (critical): a fit that ended on a non-finite log-likelihood (stop_reason nan_ll/singular_ll) round-tripped through save()/load() with no complaint, and transform() then returned all-NaN.

• state_dict() now refuses to serialize a degenerate model, and a defense-in-depth torch.isfinite scan over the param tensors blocks any non-finite parameter even if stop_reason bookkeeping misses it.
• Malformed payloads (missing wrapper/backend/config/params/extra sections or param keys) raise descriptive ValueErrors naming the file/section instead of a bare KeyError.
• _load_params guards A/comp_list shape drift vs the rebuilt config.

Code-review (important): state_dict() used .detach().cpu(), which aliases live storage when the model is already on CPU (the default). Since fit() mutates A/mu/beta in place per iteration, a snapshot captured mid-fit would silently roll forward. Now .clone()d so the snapshot is independent (verified by test_state_dict_snapshots_not_aliases). Config coverage, annealed-vs-original param handling, device/dtype restoration, and weights_only=True safety were all confirmed correct.

Test-coverage: the round-trip only exercised 6 of 11 param tensors. Added a backend-level round-trip asserting every param tensor and every extra field with do_reject+do_newton active (so mu/alpha/beta/rho/gm and the good_idx branch are non-trivial), plus unfitted / degenerate / non-finite / snapshot guard tests.

Comments: corrected _PARAM_TENSORS scope claim, the state_dict "optimizer state" wording, and the test's lossless-restore comment.

Full torch suite: 31 passed (26 baseline + 5 new persistence tests). Ruff clean.