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

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:

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.
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.
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.

Scope extension from the PR #311 review: the Phase 3 scoring accessors reproduce the same anti-pattern and postdate this issue's filing, so adding them explicitly: model_loglik/model_probability on BOTH backends (torch core.py ~3029-3106 and the MLX ports) lack the degenerate-fit/finiteness guard and the X.ndim/X.shape[0] validation, same as the item-1/item-2 accessors already listed. Also folding in: model_probability's all-non-finite error message says "-inf log-likelihood" but the isfinite check equally fires on NaN (numerical corruption), misdiagnosing the cause -- split or reword the message when fixing item 1 here. One cross-backend PR for the lot, per .rules/backend_parity.md.

Second scope extension, from the epic #278 parity audit: AMICA_NumPy.transform() (numpy_impl/core.py ~2498-2528) has the identical degenerate-fit hole — it never checks self.converged/stop_reason, so a diverged legacy-backend fit silently returns NaN sources. The issue's fix should cover all three backends, not two. (The numpy CLI's main() does check convergence before declaring success; the library path does not.)