Impute leaf input variables instead of formula aggregates in QRF

[MaxGhenis]

Summary

The QRF currently imputes formula-level aggregates (e.g. taxable_pension_income) and then renames them to leaf inputs (e.g. taxable_private_pension_income) before storing. This loses information — all taxable pension is attributed to private pensions, all interest deductions to mortgage, etc.

Current workaround (PR #594)

_rename_imputed_to_inputs maps:

• taxable_pension_income → taxable_private_pension_income (loses public pension split)
• tax_exempt_pension_income → tax_exempt_private_pension_income (same)
• interest_deduction → deductible_mortgage_interest (loses non-mortgage interest)
• self_employed_pension_contribution_ald → _person (entity mapping only)
• self_employed_health_insurance_ald → _person (entity mapping only)

Proper fix

Train the QRF on leaf input variables from the PUF rather than formula aggregates. This would:

1. Preserve the public/private pension split
2. Preserve mortgage vs non-mortgage interest
3. Eliminate the need for post-hoc renaming
4. Give more accurate distributions for each sub-component

Variables to split

• taxable_pension_income → taxable_public_pension_income + taxable_private_pension_income
• tax_exempt_pension_income → tax_exempt_public_pension_income + tax_exempt_private_pension_income
• interest_deduction → deductible_mortgage_interest + non_mortgage_interest (or deeper: investment_interest_expense)

Requires checking which sub-components are available in the PUF training data.