Refactor model parameters in the low-level interface

[andrewherren]

Description

Right now, the "low-level" stochtree interface requires a long-list of parameters for initialization and sampling, for example:

forest_model_mean <- createForestModel(
    forest_dataset, feature_types, num_trees, nrow(X), 
    alpha, beta, min_samples_leaf, max_depth
)
forest_model$sample_one_iteration(
    forest_dataset, outcome, forest_samples, active_forest, rng, 
    feature_types, leaf_model, current_leaf_scale, variable_weights, 
    a_forest, b_forest, current_sigma2, cutpoint_grid_size, 
    keep_forest, gfr, pre_initialized
)

Even if we handle this for users with good defaults, these parameters have to cover every possible leaf / variance model we support. As stochtree grows, this interface will become more and more complicated and the function calls will become lengthier.

Our proposal is to replace the long list of parameters like leaf_model and a_forest with a ModelParams or ModelConfig class which is initialized and configurable by users and then passed to functions like createForestModel() and methods like forest_model$sample_one_iteration().

Ideas for Implementation / Expected Behavior

R6 class implementing a generic interface for specifying a model and its parameters

ModelConfig <- R6::R6Class(
    classname = "ModelConfig",
    cloneable = FALSE,
    public = list(
        
        #' @field leaf_model Integer coded leaf model type
        leaf_model = NULL,
        
        #' @field feature_types Vector of integer-coded feature types
        feature_types = NULL,

        # ....
        
        initialize = function(model_type) {
            self$leaf_model <- model_type
        }, 
        
        set_variance_forest_params = function(shape, scale) {
            self$variance_forest_shape = shape
            self$variance_forest_scale = scale
        }, 
        
        # ...
    )
)

[andrewherren]

Completed in #135