Add Bayesian and sparse-selection regression (BayesianRidge, ARD, LARS)

[sipemu]

Three estimators landed in anofox-regression 0.5.6 that cover the Bayesian / sparse-selection family. None is exposed by the extension today.

Scope

Method	Rust path	Description
Bayesian Ridge	anofox_regression::solvers::BayesianRidge	Bayesian linear regression with empirical-Bayes optimization of noise (σ) and regularization (λ) via evidence maximization. Gives posterior predictive variance for free — useful for prediction intervals without bootstrap.
ARD	anofox_regression::solvers::ArdRegression	Automatic Relevance Determination — per-feature precision priors driving redundant features toward zero. Sparse feature selection for high-dimensional regressors.
LARS	anofox_regression::solvers::LarsRegressor	Least Angle Regression — efficient stepwise solver that yields the full Lasso regularization path in one fit.

Suggested user-facing names

• bayesian_ridge_fit, bayesian_ridge_fit_agg, bayesian_ridge_fit_predict_by
• ard_fit, ard_fit_agg, ard_fit_predict_by
• lars_fit, lars_fit_agg, lars_fit_predict_by — consider exposing the regularization path via an extra returned struct/list column.

Prerequisite

Bump Cargo.toml: anofox-regression = "0.5.6".

Tasks

• Bump dependency.
• FFI for each estimator; for Bayesian methods include posterior predictive variance so *_predict can populate yhat_lower / yhat_upper without bootstrap.
• C++ aggregate/table/macro wrappers.
• Decide how to expose the LARS path (single fit returning all knots, or *_path table function).
• SQL tests including a high-dim sparse-truth fixture for ARD/LARS, and a noise-tunable fixture for BayesianRidge that validates the posterior variance estimate.
• README updates.