Leverage pruning/binning for learning GAM ShapeFunctions

[jotterbach]

We can use a generalization of MDLP or the ConditionalInferenceBinner to limit the sensible depth of a RegressionTree and hence can avoid overfitting of a GAM