Generalize variance decomposition pipeline

[bridgetilin]

Multi-facet G-theory support

Extends the G-theory module from a fixed two-facet (subject × item) design to support fully crossed, balanced N-facet designs.

Changes

_generalizability.py (new) — Private helpers for multi-facet designs: design validation, interaction parsing, cell means construction, ANOVA method-of-moments variance component estimation via Möbius inversion, G-coefficient computation, and a shared bootstrap loop for both two-facet and multi-facet paths.

generalizability.py (modified) — Public API additions:

• variance_components(): multi-facet path via facet_cols, object_facet, and interactions params.
  Returns components keyed by facet name or interaction tuple, plus raw_components with unclamped estimates. Raises TypeError if both two-facet and multi-facet params are mixed.
• g_coefficient(): added facet_sizes kwarg for multi-facet mode. n_reps now defaults to observed value in multi-facet mode.
• d_study(): added design_grid and n_reps_grid_multi for multi-facet projections.
• bootstrap_variance_components(): supports both modes via shared _bootstrap_resample().

test_generalizability_multifacet.py (new) — 33 tests: variance component estimation, design validation, G-coefficients, D-study, bootstrap, ICC rejection for multi-facet dicts, and an end-to-end pipeline test.

Design notes

• Two-facet API is unchanged — existing subject_col/item_col/trial_col code works as before (42
  existing tests pass).
• Negative higher-order estimates are propagated unclamped during the top-down EMS solve to avoid
  biasing lower-order estimates. Final output is clamped to ≥ 0, and both clamped and raw values are
  returned.
• When n_reps=1, residual is confounded with the full k-way interaction; only those two are marked
  identifiable: False.
• Facet names are not hardcoded — any string column names work.

[bridgetilin]

@nicchiou