Add XGBTClassifier and XGBRClassifier meta-learners (#824)

[jeongyoonlee]

Proposed changes

Add XGBoost-backed classifier variants of the T- and R-learners — XGBTClassifier and XGBRClassifier — mirroring the existing XGBTRegressor / XGBRRegressor, and re-export both from causalml.inference.meta.

• XGBTClassifier(BaseTClassifier) — builds XGBClassifier outcome models for control and each treatment group.
• XGBRClassifier(BaseRClassifier) — uses an XGBClassifier outcome model (its predict_proba drives the outcome cross-fit) and an XGBRegressor effect model. It initializes through the grandparent BaseRLearner to bypass BaseRClassifier's both-learners-None guard, since the learners are constructed in fit().

Both follow the post-#912 scikit-learn BaseEstimator convention: constructor arguments are stored verbatim and the XGBoost models are constructed in fit(), so get_params() / clone() round-trip correctly. XGBoost hyperparameters are passed as dicts (xgb_kwargs for the T-learner; outcome_xgb_kwargs / effect_xgb_kwargs for the R-learner) and forwarded verbatim — no *args / **kwargs in __init__, which would break clone().

Fixes #824.

Credit

Reimplements the work of @jbbqqf in #899, which is CLA-blocked and cannot be merged directly. Full credit to @jbbqqf for the original implementation and analysis; this PR re-lands the feature against current master under a CLA-clean author, adapted to the BaseEstimator refactor (#912) that merged since.

Types of changes

• New feature (non-breaking change which adds functionality)

Tests

• test_XGBTClassifier / test_XGBRClassifier — fit on the synthetic classification data, assert the underlying model types (XGBClassifier outcome, XGBRegressor effect), that xgb_kwargs are forwarded, and normalized AUUC > 0.5.
• test_xgbt_classifier_clone / test_xgbr_classifier_clone — clone() / get_params() round-trip returning an unfitted estimator.
• Both classes added to _CLASSIFIER_CONFIGS, so the shared test_fit_returns_self_classifier and test_clone_get_params_classifier cover them too.
• Full tests/test_meta_learners.py — 74 passed. The new tests fail on master (classes don't exist).