Add posterior CRPS, Brier scores and finite-action risk metrics - #9

Summary

Add a reusable skbel.metrics module for:

• marginal CRPS, with weighted empirical and unweighted iid U estimators;
• multiclass Brier scores (sum across classes, explicitly documented);
• posterior expected losses over a finite action set;
• all exactly tied minimum-risk actions.

Inputs have explicit shape, finite-value and weight contracts with no implicit
broadcasting. Weight normalization rescales before summation to avoid overflow
from large finite weights. Include API notes and a small executable example.

Verification and limits

48 focused tests cover hand-computed and brute-force score comparisons, scaling
invariance, large/tiny/zero weights, invalid inputs, tied actions and ordinary
package imports. Full local upstream tests and Ruff checks pass; repository CI
checks supported Python/OS combinations. No dependency changes.

These are general scoring and terminal-decision arithmetic, not an acquisition
policy or proof of posterior calibration. CRPS is marginal, not a joint score;
the U estimator requires iid unweighted draws; Brier's two-class result is twice
the scalar binary convention. Future-observation simulation/updating and domain
loss choices remain caller responsibilities.