Add PSIS k-hat diagnostic for variational inference

[michaelellis003]

Summary

Implements the Pareto Smoothed Importance Sampling (PSIS) k-hat diagnostic for evaluating variational approximation quality, as requested in #1804.

• psis_diagnostic(rng_key, param_map, model, guide, *args) computes the k-hat statistic by fitting a Generalized Pareto Distribution to the upper tail of importance weights
• k < 0.5 indicates a good guide, 0.5–0.7 is marginal, k >= 0.7 is unreliable
• GPD fitting uses Zhang & Stephens (2009) with prior regularization from Vehtari et al. (2024), matching both Pyro's psis_diagnostic and Vehtari's reference implementation to ~1e-15

Changes

• numpyro/infer/importance.py (new): GPD fitting (_fit_generalized_pareto), PSIS tail extraction (_psis_khat), and public API (psis_diagnostic) with batched evaluation via chunk_size
• test/infer/test_importance.py (new): 36 tests covering GPD parameter recovery, cross-implementation reference values (precomputed from Vehtari's gpdfitnew and Pyro 1.9.1), regime classification against paper thresholds, batching correctness, edge cases, and SVI integration
• numpyro/infer/__init__.py: Export psis_diagnostic
• docs/source/utilities.rst: Add API documentation entry

References

• Yao, Y., Vehtari, A., Simpson, D., and Gelman, A. (2018). Yes, but Did It Work?: Evaluating Variational Inference. ICML.
• Vehtari, A., Simpson, D., Gelman, A., Yao, Y., and Gabry, J. (2024). Pareto smoothed importance sampling. JMLR, 25(72):1-58.

Fixes #1804