Improve discreteness handling, allow binary outcomes

[fverac]

Adds a binary_outcome keyword arg to most estimators, where if True then the outcome nuisance model will be a classifier.

[kbattocchi]

Mostly looks great. I've suggested a few minor changes.

Related to this PR, I think lines 52-57 of econml/dml/dml.py should be made more robust - what happens if the target is discrete and there is no predict_proba method, or target is not discrete and there is a predict_proba method? I think in either case, we should at least warn the user that they are passing a classifier where a regressor is expected or vice versa. When there is no predict_proba method but one is expected, there doesn't seem to be much harm in falling back to calling predict instead as long as the user is warned; in the opposite scenario it's less clear to me that calling predict_proba instead of just calling predict as usual is a good idea, but at least if we warn the user they can change the discreteness and get that behavior if they want.

Whatever we decide here, if there's a non-trivial amount of logic we do something similar with other estimators that don't use this first stage wrapper; probably that should happen in a new utility method that this module and others can all use consistently (perhaps something like get_prediction(estimator, expected_discrete)).

[fverac]

Mostly looks great. I've suggested a few minor changes.

Related to this PR, I think lines 52-57 of econml/dml/dml.py should be made more robust - what happens if the target is discrete and there is no predict_proba method, or target is not discrete and there is a predict_proba method? I think in either case, we should at least warn the user that they are passing a classifier where a regressor is expected or vice versa. When there is no predict_proba method but one is expected, there doesn't seem to be much harm in falling back to calling predict instead as long as the user is warned; in the opposite scenario it's less clear to me that calling predict_proba instead of just calling predict as usual is a good idea, but at least if we warn the user they can change the discreteness and get that behavior if they want.

Whatever we decide here, if there's a non-trivial amount of logic we do something similar with other estimators that don't use this first stage wrapper; probably that should happen in a new utility method that this module and others can all use consistently (perhaps something like get_prediction(estimator, expected_discrete)).

Adding a warning when first stage is discrete target but model does not have predict_proba,
and raising an error when first stage target is continuous but model does have predict_proba.

[kbattocchi]

This looks good; I've made a couple of minor suggestion but you can merge as soon as you are comfortable with it.

[kbattocchi]

I think you're including the classes themselves rather than their count here.

[fverac]

Good catch

[kbattocchi]

[minor] The singular for strata is stratum, so this name seems slightly weird. I think strata_from_discrete_arrays (since this gets the strata for all of the rows at once) is shorter and just as clear.