Use custom propensity_learner on R-learner's first fit

[arpitjain099]

Proposed changes

The R-learner ignores a user-supplied propensity_learner on the first fit() of a fresh object. In each fit() override the line self.model_p = self.propensity_learner runs after _set_propensity_models(), but _set_propensity_models reads self.model_p (via hasattr) to decide whether to use the custom propensity model or fall back to the default ElasticNetPropensityModel. On a new object model_p doesn't exist yet, so the first fit quietly uses the default and the custom learner is dropped.

Because the attribute persists on the object, a second fit() of the same instance does pick up the custom learner, so the same call gives different nuisance estimates depending on fit order. The fix moves the model_p assignment ahead of the propensity-model fitting in all three affected paths (BaseRLearner, BaseRClassifier, XGBRRegressor).

Types of changes

• Bugfix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation Update (if none of the other choices apply)

Test plan

• Added test_BaseRLearner_uses_custom_propensity_learner_on_first_fit in tests/test_meta_learners.py: it fits a fresh R-learner with a distinctive ElasticNetPropensityModel subclass and asserts the trained propensity_model is that subclass. It fails on master (default ElasticNet is used) and passes with this change.
• The R-learner test subset (-k "RLearner or RRegressor or RClassifier", 21 tests) passes locally.

Checklist

• I have read the CONTRIBUTING doc
• I have signed the CLA
• Lint and unit tests pass locally with my changes
• I have added tests that prove my fix is effective or that my feature works
• I have added necessary documentation (if appropriate)
• Any dependent changes have been merged and published in downstream modules

[jeongyoonlee]

LGTM