Add propensity clipping diagnostics to DRTester

[leninworld]

This PR fixes issue #1054.

Summary

• make the lower propensity clipping threshold configurable in calculate_dr_outcomes and DRTester
• report structured train/validation clipping diagnostics by observed treatment arm
• warn only when clipping changes a doubly robust denominator, while preserving all-row overlap counts
• validate non-finite inputs, incompatible shapes, and missing control treatment 0

Closes #1054.

Validation

• python -m pytest econml/tests/test_validate_utils.py econml/tests/test_drtester.py -q -o addopts='' -p no:cacheprovider — 11 passed, 6 subtests passed
• Ruff lint on all changed files — passed
• Python compilation and git diff --check — passed
• Existing real-model DRTester tests exercise binary, multi-treatment, train/validation, and dataframe paths

Validation evidence

Compatibility and scope

The default threshold remains 0.01, and default arithmetic is unchanged when no clipping occurs. The new warnings distinguish clipping that affects an assigned treatment denominator from low propensity values on unassigned rows.

This PR intentionally does not implement trimming because trimming changes the analyzed sample and can change the estimand. Upper clipping and propensity calibration are also out of scope.

The validation now raises explicit ValueError exceptions for non-finite inputs and for data without control treatment 0 instead of allowing invalid doubly robust output.