Fix bootstrap inference

[kbattocchi]

No description provided.

[vsyrgkanis]

Currently we have no test that calls coef__inference or summary() but such that the estimator has no cate_feature_names method!

We either need to add such a test and make sure the behavior is as we want:

1. I see this here being problematic because we would ideally want when I call summary(feat_name=..), even if the method has no cate_feature_names method, that the feat_name list that I give, will appear in the summary table.

Or maybe even better: we make cate_feature_names a mandatory method for any estimator that inherits from LinearModelFinalInference class and the LinearModelFinalInferenceDiscrete class. In this case, summary should perform ok as we will always have cate_feature_names implemented, whenever summary is available.

[vsyrgkanis]

We should add a test that checks that the attribute error is raised in this loop and nothing else happens.

[vsyrgkanis]

Here T takes three values {0, 1, 2}

est2 = LinearDRLearner(model_regression=RandomForestRegressor(n_estimators=10),
                       model_propensity=RandomForestClassifier(n_estimators=10, min_samples_leaf=50),
                 featurizer=None, min_propensity=1e-3,
                 n_splits=6)
est2.fit(Y, T, X, W, inference=BootstrapInference(n_bootstrap_samples=8))
est2.coef__inference(T=1).summary_frame()

Should be returning the inference only for the coefficient associated with T=1. Currently it returns inference for all models.

Moreover, est2.summary(T=1) raises an error. Weird since, I think we do have a test for summary with drlearner and bootstrap.