Make prediction methods more flexible

[andrewherren]

Right now predict.bartmodel and predict.bcfmodel in R, and the predict() methods in Python return a list / dictionary of posterior predictions for each model term.

Since a common workflow is to extract the posterior mean prediction for either $\hat{y}$ in BART or $\hat{\tau}$ in BCF, we might consider rewriting these prediction routines to allow a type = 'mean' style argument that avoids making users extract the posterior matrix for a given model term and then take its row-wise average.