Add intraclass correlation coefficient (ICC) to generalizability metrics

[sonnetx]

Summary

Adds intraclass_correlation() to src/torch_measure/metrics/generalizability.py. ICC is a ratio of the variance components already estimated by variance_components (subjects = targets, items = raters), so it slots in next to g_coefficient and reuses the same dict.

Supported forms (McGraw & Wong 1996 / Shrout & Fleiss 1979 two-way):

• ICC2 — two-way random, single rater, absolute agreement
• ICC3 — two-way mixed, single rater, consistency
• ICC2k — average of k raters, absolute agreement
• ICC3k — average of k raters, consistency

The averaged forms are mathematically identical to the existing G-coefficient at n_reps=1 (ICC2k == g_coefficient(type="absolute"), ICC3k == g_coefficient(type="relative")) which is used as the primary correctness test. One-way forms (ICC1/ICC1k) require a different (one-way) decomposition and raise a clear ValueError rather than returning a wrong number.

Test plan

• pytest tests/test_metrics/test_generalizability.py — 33 tests pass (10 new in TestIntraclassCorrelation: unit-interval, G-coefficient equivalence, consistency ≥ absolute, average ≥ single, n_items defaulting, ICC1/unknown-form/missing-keys errors, zero-components, real-components)
• pytest tests/test_metrics/ — 74 pass, no regressions
• ruff check / ruff format --check — clean

Note for reviewers

This stacks on #27 (G-theory base). Once #27 lands I'll rebase onto main and the diff will collapse to just the ICC commit, then flip this out of draft.

First of several planned G-theory follow-ups (bootstrap CIs, REML backend, assumption checks, a walk-through tutorial).