Add joint posterior scoring and finite adaptive acquisition

[robinthibaut]

Summary

• Ignore zero-probability draws in weighted CRPS calculations.
• Add an empirical joint energy score with strict shape and numerical validation.
• Add finite adaptive acquisition and stopping over supplied joint atoms, with horizons zero through two, explicit costs and immutable policy trees.
• Preserve the existing ranking exports and document the new APIs and their limits.

Validation

All 267 tests pass, including independent pairwise and rational-policy oracles and an identical-marginal dependence control. Ruff lint and formatting checks pass.

These APIs evaluate the supplied finite distributions. They do not establish calibration, choose coordinate scales or supply an application-specific loss model. Floating-point representability and combinatorial complexity limits are documented.