Move the training loop boilerplate code

[jason-fries]

Currently the training loop is in ehr_ml/blob/master/ehr_ml/clmbr/__init__.py. Training should be external to the specific model choice to better support other architecture choices and hook into existing training/tunining frameworks like Ray Tune, pytorch lightning

[woffett]

I moved the training loop to ehr_ml/clmbr/featurizer inside a .fit method. Not exactly external to the model (actually kind of built-in to the featurizer class now) but in the featurizer class does build the model based on the config, so in that sense this kind of accomplishes the goal of model-agnostic training code?