Blend SIPP and SCF financial asset imputations

[MaxGhenis]

Summary

• applies a stable household-level 50/50 source-model draw between SIPP and SCF for overlapping bank, stock, bond, and vehicle-value assets
• fixes the CPS builder to blend from the persisted full CPS/SIPP asset arrays before computing formula net_worth, so masked SCF receiver rows do not drop overlapping assets
• adds SCF-only balance-sheet leaves for certificates of deposit, savings bonds, retirement assets, insurance, managed assets, real estate, business equity, vehicle/student/other installment debt, mortgages, credit cards, and other debts
• computes net_worth directly from the final mixed-source component formula, with no net_worth_residual; the internal direct SCF net-worth prediction anchors proportional rebalancing of SCF-only leaves while preserving resource-tested policy leaves
• validates that the SCF source provides every formula target needed for the component sum
• adds repo-level AGENTS.md instructions so future agent PRs use in-repo PolicyEngine branches rather than forks
• updates docs, changelog, and tests for the component formula and household-level source draw

Refs #835.

Validation

• uv run pytest tests/unit/test_asset_imputation.py tests/unit/calibration/test_source_impute.py tests/unit/datasets/test_cps_file_handles.py
• uv run ruff format --check .
• uv run ruff check policyengine_us_data/utils/asset_imputation.py policyengine_us_data/datasets/cps/cps.py policyengine_us_data/calibration/source_impute.py tests/unit/test_asset_imputation.py tests/unit/calibration/test_source_impute.py tests/unit/datasets/test_cps_file_handles.py
• uv run towncrier check --compare-with origin/main
• git diff --check