Add G-theory modeling-assumption checks (normality, Tukey nonadditivity)

[sonnetx]

Follow-up to #26. Stacks on #27, #35, #36.

Summary

Adds a new src/torch_measure/metrics/assumptions.py module with two diagnostics for the G-theory variance-component model:

• normality_check(response_matrix, target=...) — Shapiro-Wilk on a chosen effect family ("subject", "item", "subject_item", or "residual"), plus the sorted sample and theoretical quantiles needed for a Q-Q plot.
• tukey_nonadditivity_test(response_matrix) — Tukey's 1-df test on cell means. Especially load-bearing when n_reps=1, where subject_item and residual are confounded in variance_components and this test asks whether the combined residual contains structured interaction beyond noise.

Both functions return the test statistic, p-value, and the underlying numbers (Q-Q quantiles for normality, SS components for Tukey). Levene (homogeneity) and balance/missingness checks deferred but happy to add those or other checks as needed.

Test plan

• 14 new tests in tests/test_metrics/test_assumptions.py:
  • Normality: structure, normal residuals not rejected, heavily-skewed (lognormal) residuals rejected, each target variant works, target="residual" requires replications, validation errors.
  • Tukey: structure, additive synthesis not rejected, multiplicative synthesis rejected (with and without replications), unbalanced design raises.
• Full tests/test_metrics/ suite green (97 tests)
• ruff check / ruff format clean

[nicchiou]

Please include Levene's test for homogeneity of variance to this PR as well.

[sonnetx]

Added levene_homogeneity_test in d046e6d. Defaults to Brown-Forsythe (median-centered) for robustness; supports group_by="item" (default) or "subject" and returns the per-group sizes and variances alongside the statistic so callers can spot small-group issues. 7 new tests; full assumptions suite is 22 green.