Donor-impute CPS demographic and occupation features on PUF clones

[MaxGhenis]

Summary

• donor-rematch clone-half is_male, cps_race, is_hispanic, and detailed_occupation_recode using CPS records with similar demographics, geography, and PUF-imputed incomes
• rebuild the overtime occupation input flags from the imputed occupation code so FLSA overtime logic uses coherent clone-half inputs
• make stage-2 CPS-only QRF predictors read from the clone half itself when those person-level features have been updated
• replace star imports in extended_cps.py with explicit dataset imports and add focused unit coverage for clone-half feature rematching

Why

Right now the PUF clone half keeps donor-copied CPS demographic and occupation labels even after its income profile is replaced by PUF imputation. That weakens subgroup analysis by race/ethnicity/sex and can misclassify occupation-based logic like overtime exemptions.

This PR keeps the existing stage-2 QRF for continuous CPS-only variables, but adds a preceding donor-imputation pass for the categorical CPS features we actually use downstream.

Testing

• uv run ruff check policyengine_us_data/datasets/cps/extended_cps.py policyengine_us_data/tests/test_extended_cps.py
• uv run pytest -q policyengine_us_data/tests/test_extended_cps.py policyengine_us_data/tests/test_calibration/test_puf_impute.py

Interaction notes

• #633 also touches policyengine_us_data/datasets/cps/extended_cps.py, but only for structural mortgage input support. I kept this PR based on main because the logic is otherwise independent; the overlap should be limited to import/generate() context when that PR merges.
• #631 is broader pipeline restructuring and does not change the clone-feature logic added here.

[MaxGhenis]

Refreshing CI after moving this PR from a fork branch to an upstream branch; reopening immediately.

[MaxGhenis]

Superseded by #658, which uses a fresh upstream branch for the same change.