Fix MVN noise handling and train-only CompositePCA scaling

[robinthibaut]

Summary

• Assign and validate explicit BEL.predict(noise=...) values before the MVN
  covariance path consumes them. Preserve the historical None reset to0.01.
• Fit CompositePCA's optional scalers on training PCA scores once, reuse them
  during transform and undo them before inverse projection.
• Restore batched inverse transforms and derive score widths from fitted PCA
  components, including None and variance-fraction configurations.

The noise parameter retains its existing projected-covariance convention; this
does not reinterpret it as physical-space observation standard deviation.

Verification

Eight new small regression tests capture actual MVN covariance inputs, rejected
noise values, held-out batch invariance, frozen training scalers, roundtrips,
component settings and shape/fittedness errors. The full local upstream suite
and Ruff checks pass; the repository's CI matrix provides additional Python/OS
compatibility checks. No dependencies or other inference algorithms change.

This fixes software contracts. It does not establish posterior calibration or
application-specific scientific performance. Refit previously fitted scaled
CompositePCA objects to establish the new training-scaler state.