Improving on access to first stage nuisance models

[vasilismsr]

Currently in DMLCateEstimators we return the scikit-learn models_y and models_t and not the first stage wrappers around them. This makes it hard for someone to access these models and predict on a set of functions to see how well they performed, because they would need to combine the X and W themselves. We mostly did this so that the user can access easily the properties of the first stage models as if it was a scikit-learn estimator. But this makes prediction hard.

One solution would be to return the first stage wrappers, but we need to make the first stage wrappers inherit the properties of the wrapped models, so that the user can do:
model.predict(X, W)
but then also be able to do things like
model.coef_
model.feature_importances_
etc.