Use Cholesky solves in PMC covariance evaluation

[fujikix1102]

Summary

• replace explicit covariance inverses in PMC Gaussian and Student-t components with cached Cholesky factors
• evaluate quadratic forms with triangular solves and compute log normalizations from the Cholesky diagonal
• retain the historical A normalization attribute for compatibility
• add regression coverage against the direct quadratic-form reference

Validation

• pytest -q cosmosis/test/test_samplers.py -k 'pmc or pmaxlike or maxlike' --disable-warnings --maxfail=1 (17 passed)
• Python compilation and git diff --check passed

The complete sampler test file is environment-limited here because the optional dynesty dependency is not installed.

[fujikix1102]

Follow-up validation: installed dynesty, zeus-mcmc, and nautilus-sampler in the test environment. Full cosmosis/test/test_samplers.py now passes: 37 passed, 4 skipped, 457 warnings in 46.71s. The skips are optional/unsupported environments; no PMC-related failures.

[fujikix1102]

Performance/reproducibility check: with the same seed (20260924), dimension=24, 4,000 samples, and 8 log-density evaluations per timing run, the Cholesky implementation had median 0.00335 s versus 0.03759 s for the previous explicit-inverse implementation (11.2x speedup). Maximum absolute log-density difference was 1.42e-13.