Expose CATE estimation and inference methods on DRPolicyForest/Tree

[jcharit1]

Summary

• DRPolicyForest and DRPolicyTree internally fit a DRLearner but do not expose its CATE estimation and inference API (e.g., effect(), effect_inference(), effect_interval(), shap_values(), score())
• Users who need both optimal policy assignments and CATE confidence intervals may consider strategies such as a 3-way data split, which can be less data-efficient
• EconML's cross-fitting within DRPolicyForest already provides noise separation between CATE estimation and policy learning — exposing the underlying CATE inference methods enables a more data-efficient 2-split workflow

Changes

• Add delegation methods on _BaseDRPolicyLearner for effect(), effect_interval(), effect_inference(), const_marginal_effect(), const_marginal_effect_interval(), const_marginal_effect_inference(), marginal_effect(), ate(), ate_interval(), ate_inference(), shap_values(), score(), and model_final_
• Pass inference parameter through _BaseDRPolicyLearner.fit() to the underlying DRLearner
• Fit per-treatment RegressionForest CATE models alongside the policy model in _PolicyModelFinal to support GenericModelFinalInferenceDiscrete
• Override _get_inference_options in _DRLearnerWrapper to enable automatic inference when CATE models are available

Motivation

When learning personalization policies from A/B test data, a natural workflow is:

1. Learn optimal treatment assignments using DRPolicyForest
2. Estimate CATEs with confidence intervals for the policy-assigned treatments

Without this change, users must manually split data across separate ForestDRLearner and DRPolicyForest fits, losing the data efficiency that EconML's built-in cross-fitting provides.

Test plan

• All 10 existing test_policy_forest.py tests pass
• All 234 non-Ray tests in the full suite pass (0 regressions)
• Verified effect(), effect_inference(), effect_interval(), const_marginal_effect_inference(), shap_values(), score() all work on DRPolicyForest after the change
• Verified inference produces valid point estimates, standard errors, p-values, and confidence intervals

🤖 Generated with Claude Code