Handle overlapping regime-aware donor predictors

[MaxGhenis]

Summary

• remove target variables from the base predictor list before fitting the chained regime-aware donor imputer
• preserve sequential chaining so earlier targets are still used as predictors for later targets
• add a regression test for the production failure mode where a target also appears in the original condition vars

Failure fixed

The post-#241 release rebuild failed in the large PUF chained income block because duplicate column names caused pandas to return a DataFrame for a target, and MicroImpute then crashed in target-type detection on .unique().

Tests

• env -u UV_FROZEN uv run --python 3.13 --extra dev --extra policyengine ruff format src/microplex_us/pipelines/donor_imputers.py tests/pipelines/test_regime_aware_donor_imputer.py
• env -u UV_FROZEN uv run --python 3.13 --extra dev --extra policyengine ruff check src/microplex_us/pipelines/donor_imputers.py tests/pipelines/test_regime_aware_donor_imputer.py
• env -u UV_FROZEN uv run --python 3.13 --extra dev --extra policyengine pytest -q tests/pipelines/test_regime_aware_donor_imputer.py
• env -u UV_FROZEN uv run --python 3.13 --extra dev --extra policyengine pytest -q tests/test_pe_source_impute_specs.py tests/test_pe_source_impute_engine.py tests/pipelines/test_donor_imputer_negative_preservation.py tests/data_sources/test_scf_net_worth_components.py -k "imputer or regime_aware or columnwise or pe_source"