Update anofox-regression to 0.5.2 with new features

[sipemu]

Summary

• Crate update: anofox-regression 0.5.1 → 0.5.2
• SolverType selection: QR (default), SVD (most robust), Cholesky (fastest) for OLS, Ridge, WLS
• LambdaScaling: Raw (as-is) and Glmnet (R's glmnet convention) for Ridge and ElasticNet
• Penalized IRLS: L2 regularization (lambda) parameter for all GLM models (Poisson, Binomial, NegBinomial, Tweedie)
• HC inference: Heteroscedasticity-consistent standard errors (HC0–HC3) for OLS and WLS
• Native WlsRegressor: Replaced manual sqrt(w) transformation with upstream WlsRegressor for proper weighted statistics
• LmDynamic: Time-varying coefficient model using IC-based model averaging with optional LOWESS smoothing
• Condition diagnostics: Condition number and severity assessment for design matrices
• Quasi-separation detection: Binary separation and count sparsity checks for GLM data
• LOWESS smoothing: Standalone locally weighted scatterplot smoothing function
• FFI wiring: All new features wired through C-compatible FFI layer

Test plan

• All 144 unit tests pass (was 110 before, added 34 new tests)
• Zero compiler warnings
• Tests cover: solver selection (SVD, Cholesky), lambda scaling, penalized GLM, HC inference (HC1, HC3), native WLS, LmDynamic (basic, no smoothing, BIC, edge cases), condition diagnostics, separation detection, LOWESS smoothing

🤖 Generated with Claude Code