Expanding validation options for meta-learners

[bsaunders27]

Still getting up to speed on the literature around CATE, but noticed that the docs on validation didn't have anything problem specific, and I wondered if there were some case-specific validations we could perform at least on the fitness of base-learners (mu and treatment propensity). It seems like it'd be fairly straightforward to implement a .score_baselearners(X,Y,T) method which could return metrics that could detect whether the base-learners were over/under-fitting.

Currently accessible via the .models_mu_(t|c) but thinking this might have value to add as a dedicated method for model evaluation. Happy to take a stab at implementing if folks think this would be useful/worth adding as a native feature.

from sklearn.metrics import roc_auc_score
xlearner = BaseXClassifier(outcome_learner=XGBClassifier(),
                           effect_learner=XGBRegressor())
xlearner.fit(train[covariates], train['treatment'], train['target'])
xlearner.models_mu_t[1].score(
    test.loc[test['treatment']==1, covariates],
    test.loc[test['treatment']==1, 'target'])

roc_auc_score(train.loc[train['treatment']==1, 'target'],
              xlearner.models_mu_t[1].predict_proba(train.loc[train['treatment']==1, covariates])[:,1]
)