Align itemized deduction targets with SOI itemizers and JCT tax expenditures

[baogorek]

Summary

Fixes #646.

This PR does two related things for itemized-deduction calibration targets:

• aligns IRS SOI itemized-deduction targets to the itemizing-filer population
• aligns JCT national deduction/QBI targets to their correct economic meaning as tax expenditures

Changes:

• add tax_unit_itemizes == 1 to the SOI strata for salt, real_estate_taxes, and medical_expense_deduction
• load national JCT deduction and QBI rows as reform-based tax expenditure targets instead of baseline filer/itemizer targets
• carry reform_id through target_overview, target querying, matrix building, and staging validation
• compare reform targets to repeal-based income_tax deltas rather than raw deduction totals
• deactivate stale baseline rows from older ETL runs and ignore inactive rows in matrix target selection
• add regression coverage for reform_id passthrough and reform target querying/naming

Why

The original SOI fix was directionally correct, but the JCT side still treated tax expenditures as deduction aggregates. JCT publishes revenue-loss estimates, not raw deduction totals, so calibration needs to target repeal-based income_tax deltas for those rows.

Validation

• python3 -m py_compile on all edited files
• SQLite smoke test of the updated target_overview and best-period target query logic

Environment limitation

I could not run the repo's normal pytest path in this workspace because the available environments are broken on torch and missing sqlmodel in the fallback Python.

[MaxGhenis]

Reviewed the economics and the calibration-path changes in detail.

The SOI-side itemizer constraint fix is correct, and the follow-up JCT changes now fix the real concept mismatch: these national JCT rows are tax expenditures, not baseline deduction aggregates. With the reform_id plumbing in place, the PR is now targeting repeal-based income_tax deltas rather than raw deduction sums, which is the right object.

I don't see a remaining blocking correctness bug. The follow-up items I still want are non-blocking cleanup work:

• simplify the validate_staging.py reform cache/helper split so it can't be misused later
• add tests for stale-row deactivation in etl_national_targets.py
• add a test around reform-delta computation / sign convention
• optionally add explicit source comments for the hardcoded JCT table values

I'm planning to merge this and open a focused cleanup PR for those items.