Update BART prediction functions to reflect cloglog

[andrewherren]

Right now, the predict function / method in both R and Python contains a scale argument whose valid options are 'linear' and 'probability'. We will modify this to more closely match the spirit / style of existing GLM implementations. For a quick overview:

• the glm function in base R uses type = 'response' for inverse-link scale and type = 'link' for linear scale
• the GLM class in statsmodels (Python) uses which = 'mean' for inverse-link scale and which = 'linear' for linear scale

We should accept either 'response' or 'mean' for inverse-link scale and 'link' or 'linear' for linear scale. We will also continue to accept 'probability' for probit models, with a deprecation warning that it will be phased out in future versions.

We should also consider whether ordinal models are best served by additional helper functions for computing, e.g. the probability of an observation attaining a given category value. It might make sense to allow linear / link scale predictions for ordinal models and then ask users to pass the predictions through post-processing functions for computing specific probabilities.