Replace batched QRF with sequential fit_predict()

[MaxGhenis]

Summary

• Replace _batch_qrf() (groups of 10 variables, destroying covariance) with _sequential_qrf() using microimpute's fit_predict() — single sequential QRF preserving full joint distribution across all 85+ PUF income variables
• Replace manual sampling/gc in _impute_weeks_unemployed() and _impute_retirement_contributions() with max_train_samples parameter
• Bump microimpute dependency to >=1.15.1

Fixes the covariance destruction from batching (related: #590). Upstream features added in PolicyEngine/microimpute#170.

Test plan

• CI passes (unit tests with synthetic data)
• Full dataset build produces reasonable distributions
• Verify income variable correlations are higher than batched approach

🤖 Generated with Claude Code

[baogorek]

The old _batch_qrf explicitly zero-filled variables missing from the training data and included them in the returned dict. The new _sequential_qrf returns only predictions.columns, so if fit_predict() drops any requested variable, the callers at lines 511, 514, and 537 will KeyError — they do bare y_full[variable] lookups with no guard.

Rather than silently zero-filling like before, I'd suggest failing loud:

result = {var: predictions[var].values for var in predictions.columns}
missing = set(output_vars) - set(result)
if missing:
    raise ValueError(
        f"{len(missing)} variables requested but not returned by "
        f"fit_predict(): {sorted(missing)[:10]}"
    )
return result

[MaxGhenis]

Good catch — added the explicit ValueError in d2114a6. microimpute's fit_predict() does zero-fill missing vars so they'll always be in predictions.columns, but better to fail loud if something unexpected happens.

Also found that _drop_formula_variables was silently dropping 6 PUF-imputed adds variables (taxable_pension_income, interest_deduction, tax_exempt_pension_income, pre_tax_contributions, self_employed_pension_contribution_ald, self_employed_health_insurance_ald) whose sub-components aren't separately stored. Added them to _KEEP_FORMULA_VARS in the same commit.

[baogorek]

LGTM