Add information criteria for explicit-state fits

[jonathanberthias]

The explicit-state fitting path has a common log_prob objective and retains its fitting observations, but information criteria still live in the legacy metrics module, which reads mutable distribution parameters and cannot score a FitResult. The likelihood helper can also silently sum an extra model batch axis.

This PR adds aic(fit) and bic(fit) for explicit-state MLE results. Both evaluate the model at the result's state on its retained fitting dataset, count only free parameter coordinates, and require a successful fit and finite evaluated likelihood. BIC counts observations along the first data axis, with scalar data counting as one observation. Dataset likelihood aggregation requires one pointwise log score per observation.

AIC and BIC are evaluated after MLE fitting. For a fixed model their penalties are constant in the parameter values, so minimizing them would give the same fitted state as minimizing negative log likelihood. Predictive scores that can change the fit are handled in a separate slice.