Multicollinearity diagnostics + GLM χ² (#27 batch 4)

[sipemu]

Batch 4 of 4 for #27. Adds 4 multicollinearity + dispersion-related diagnostics.

Implemented

Diagnostic	Expression	Output
High-VIF mask	`ps.high_vif_predictors("x1", "x2", ..., threshold=10.0)`	struct{is_high, n_high, n_features}
Generalized VIF	`ps.generalized_vif("x1", ..., group_sizes=[2, 3])`	struct{gvif, n_groups}
Logistic Pearson χ²	`ps.pearson_chi_squared_logistic("y", "x1", ...)`	struct{chi_squared, df_resid, n_observations}
Poisson Pearson χ²	`ps.pearson_chi_squared_poisson("y", "x1", ...)`	same

All four follow the rlib split pattern. The Pearson χ² statistics are just Σ pearson_residual² — reusing batch-2 plumbing without needing the generic GlmFamily trait.

Tests

• Python — tests/test_mc_dispersion.py, 9 new pytests:
  • VIF mask flags all 3 correlated columns at threshold 5, none of 3 independent columns at threshold 10.
  • GVIF on 5 features split into groups [2, 3] returns 2 GVIF values.
  • χ²/df_resid lands within an order of magnitude of 1.0 for well-specified data on both logistic and Poisson.
• Rust — tests/rust_api.rs: 4 smoke tests appended to diagnostic_fits.

Full Python suite: 457/457 pass (448 prior + 9 new). Rust integration: 12/12 pass.

Issue #27 status after this PR

This completes the user-facing scope I had time for. Remaining items I intentionally deferred:

• estimate_dispersion_pearson / estimate_dispersion_deviance — trivially computed from χ²/df_resid, so callers can derive them from this PR's output.
• classify_condition_number — ConditionSeverity is already returned by the existing condition_number diagnostic.
• variance_decomposition_proportions (Belsley-Kuh-Welsch) — returns a full Mat, needs a more complex output schema.
• Probit / cloglog / negative binomial / Tweedie residual variants — same *_pearson_residuals / *_deviance_residuals / *_working_residuals triple for each family, ~12 more functions.
• Standardized Pearson / standardized deviance residuals — need GLM leverage, which isn't currently exposed.

Recommend closing #27 with this PR and filing a smaller follow-up issue for the GLM extensions if/when there's demand.

Test plan

• cargo build --no-default-features clean
• cargo clippy --all-targets -- -D warnings clean
• cargo fmt --check clean
• cargo test --no-default-features — 12/12 pass
• pytest — 457/457 pass
• CI matrix

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 93.33333% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
python/polars_statistics/exprs/regression.py	93.33%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!