Guard raw-backend accessors on degenerate fits and bad transform input

[neuromechanist]

From the PR #305 review (epic #278 Phase 1). Three gaps exist identically in AMICATorchNG and (by deliberate parity) AMICAMLXNG; fixing them one-backend-only would violate .rules/backend_parity.md, so they are tracked here for one cross-backend PR:

1. transform/get_mixing_matrix/get_unmixing_matrix/get_sensor_mixing_matrix return numeric-looking output from a fit the backend itself classified as degenerate (stop_reason in the degenerate set, or a non-finite parameter): no warning, no error. get_rho already carries a finiteness guard on both backends, and state_dict refuses to persist such a model, so the fix pattern exists. The AMICA wrapper's AMICA wrapper marks is_fitted_ even on a degenerate (nan_ll/singular_ll) fit #50 contract (_check_usable) already protects wrapper users; this is about the raw backend surface.

2. transform(X) does not validate X.ndim/X.shape[0] the way fit does on the same classes, so a wrong-channel-count array fails with a raw matmul/broadcast error instead of the named ValueError fit gives for the identical mistake.

3. from_state_dict's cls(**config) surfaces malformed config as a bare TypeError (unexpected/missing kwargs) while every neighboring validation step raises a named ValueError identifying the save file as the culprit.

Same-PR cross-backend fix with a shared test per .rules/backend_parity.md.