Add bootstrap CIs for variance components

[sonnetx]

Follow-up to #26. Builds on #27 (G-theory module) and #35 (ICC).

Summary

Adds bootstrap_variance_components() to src/torch_measure/metrics/generalizability.py: a nonparametric bootstrap that resamples subjects with replacement, re-fits the moments estimator on each draw, and returns percentile CIs for each variance component plus the full bootstrap distribution.

out = bootstrap_variance_components(
    response_matrix, n_boot=2000, ci=0.95, seed=42,
)
out["subject"]              # point estimate
out["ci"]["subject"]        # (lo, hi) percentile interval
out["samples"]["subject"]   # np.ndarray of n_boot draws

Subjects are the exchangeable unit; duplicates from sampling-with-replacement are relabelled so each draw is treated as a distinct unit. Degenerate draws (singular cell structure) are skipped via try/except.

The returned samples dict lets callers compose CIs for any function of the components without re-running the bootstrap — including g_coefficient and intraclass_correlation.

Test plan

• 9 new tests under TestBootstrapVarianceComponents: output structure, point-estimate pass-through, reproducibility under seed, different seeds differ, CI brackets the point estimate on a well-separated design, wider CI at higher confidence level, input validation, and end-to-end G-coefficient CI via samples
• Full module suite (42 tests) green; ruff clean

Stacking

This branch stacks on top of #35 (which stacks on #27). After #27 and #35 merge, I'll rebase onto main and mark this ready.