Add state-aware predictive scores

[jonathanberthias]

The legacy CRPS, Brier, and quantile scores read distribution values implicitly and cannot evaluate an explicit parameter state. The legacy CRPS also squares the difference from the empirical CDF, which is not the mean score of individual forecasts when observations have different predictions.

This PR adds state-aware CRPS, Brier, and quantile scores for univariate continuous models. Each takes a structural model, scalar or one-dimensional observations, and an optional explicit state, so the same function can score a candidate during optimization or inspect a fitted state. Forecasts must be scalar or aligned with observations. CRPS averages individual scores and splits quadrature at observations and finite support endpoints and medians, so narrow bounded forecasts are resolved.

AIC and BIC remain separate post-fit diagnostics because their parameter penalties are constant for a fixed model. Discrete threshold events need their own probability capability before Brier scoring is generalized beyond continuous laws.