Bound regime-aware QRF donor fits

[MaxGhenis]

Summary

• bump Microplex's policyengine extra to MicroImpute 3.1.1 for the chained-imputation feature-order fix
• propagate regime-aware QRF forest settings into the actual MicroImpute base QRF
• add a configurable QRF training-row cap and run QRF fits with n_jobs=-1 so production PUF donor blocks are bounded

Context

The post-rollback MP rebuild uncovered that the correct chained MicroImpute path was finally using prior imputations as predictors, but the production-scale regime-aware donor adapter was still handing an unbounded default QRF into MicroImpute. That made large PUF donor blocks slow and also ignored the configured n_estimators value.

Validation

• uv run ruff check src/microplex_us/pipelines/donor_imputers.py src/microplex_us/pipelines/us.py tests/pipelines/test_regime_aware_donor_imputer.py tests/pipelines/test_us.py
• env -u UV_FROZEN uv lock --check
• env -u UV_FROZEN uv run --python 3.13 --extra dev --extra policyengine python -m pytest -q tests/pipelines/test_regime_aware_donor_imputer.py tests/pipelines/test_us.py -k 'qrf_budget_reaches_microimpute_base or multi_target_fit_uses_one_chained_zero_inflated_imputer or support_sensitive_donor_vars_do_not_force_clamps'
• env -u UV_FROZEN uv run --python 3.13 --extra dev --extra policyengine python - <<'PY'
  from importlib.metadata import version
  print(version('microimpute'))
  PY