implement multi-process (single node) parallelization for training and predicting

[timodonnell]

@rohanpai has done some proof of concept work on this and has found it can be a big win when using CPU

Should be exposed the same as scikit-learn handles parallelism using joblib. For example, see the n_jobs argument here: http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html