Implements an extension of Bayesian Additive Regression Trees (BART)
in which each regression tree outputs a linear combination of random ridge functions
(i.e., a composition of a non-linear function like cosine, hyperbolic tangent, the rectified linear
unit with an affine transformation) instead of a constant.
Can be used to perform "targeted smoothing" in which trees split on certain covariates
but output smooth functions in other covariates. For more information, see
Yee, Ghosh, and Deshpande (2026+)