QRF-impute CPS-only variables for PUF clone half

[MaxGhenis]

Summary

Fixes #561

The extended CPS doubles the dataset — CPS records on the left, PUF clones on the right. Income variables are already QRF-imputed from PUF data, but ~60 other CPS-only variables (retirement distributions, transfers, hours, medical expenses, Social Security sub-components, SPM variables) were naively duplicated from the CPS donor regardless of the PUF clone's imputed income.

This PR adds a second-stage QRF imputation so the PUF clone half gets plausible values for these CPS-only variables that are consistent with its imputed income profile:

• New CPS_ONLY_IMPUTED_VARIABLES list (~60 variables): retirement distributions/contributions, Social Security sub-components, transfer income, SPM variables, medical expenses, hours/employment, and previous-year income
• New impute_cps_only_variables() function: trains a QRF on CPS person-level data where predictors = demographics (age, sex, filing status, dependents) + key income variables (employment, self-employment, Social Security), and outputs = the CPS-only variables. For PUF clone prediction, the income predictor columns are swapped with PUF-imputed values from the first stage.
• Updated concatenation loop: CPS-only imputed variables now use original CPS values for the CPS half and second-stage QRF predictions for the PUF clone half (same pattern as existing IMPUTED_VARIABLES)

Uses the same batching, memory management, and entity-mapping patterns as the existing QRF code.

Test plan

• Verify the extended CPS generates without errors (make data)
• Spot-check that PUF clone half values for CPS-only variables differ from CPS donor values and correlate with imputed income
• Confirm CPS half values are unchanged

[baogorek]

@MaxGhenis , The more I review this, the more I realize that imputation is just as complex as calibration (if not more) and probably should have its own dashboards, etc. But for now, my review is in two inline comments plus other comments from Claude below.

Other CC comments:

 SS sub-components (line 39): Same overwrite issue as the retirement contributions. reconcile_ss_subcomponents() in  
  puf_clone_dataset() ensures these 4 shares sum to 1 and match the imputed social_security total. Stage 2's
  unconstrained QRF replaces them with values that have no summing guarantee.                                         
                                                            
  Pension leaf variables (line 31): taxable_private_pension_income and tax_exempt_private_pension_income get silently 
  discarded. _rename_imputed_to_inputs() runs after Stage 2 and overwrites them with the PUF-imputed values. Wasted
  QRF effort.

  Formula variables (line 75): weekly_hours_worked and employment_income_last_year have formulas/adds in
  policyengine-us, so _drop_formula_variables() removes them after Stage 2. Predictions are wasted unless they're
  added to _KEEP_FORMULA_VARS.

  child_support_expense (line 73): Listed under # Medical expenses but it's not a medical expense.

  Test gap (test file line 33): Should also check overlap with CPS_RETIREMENT_VARIABLES and SS_SUBCOMPONENTS from
  puf_impute.py — those have specialized imputation that Stage 2 would overwrite.

[baogorek]

At first I was thinking that "cloning" was referring to the stage 2 stuff, but there are "cloning" operations here as well. CC is saying that:

• traditional_401k_contributions
• roth_401k_contributions
• roth_ira_contributions
• self_employed_pension_contributions

"are already imputed by _impute_retirement_contributions() in puf_clone_dataset() with domain-specific constraints — IRS contribution limits by year, age-based catch-up provisions, and income requirements (e.g., 401k zeroed when employment_income = 0).Stage 2's unconstrained QRF overwrites these with values that violate legal limits. "

● 1. puf_clone_dataset() imputes retirement contributions for the PUF clone half with explicit IRS constraints (e.g., 
  401k capped at $23,000 + catch-up, zeroed when no employment income)
  2. _impute_cps_only_variables() trains a generic QRF that also predicts those same 4 retirement variables — with no 
  constraints                                               
  3. _splice_cps_only_predictions() overwrites the constrained values from step 1 with the unconstrained values from  
  step 2                                                    

Okay, and later it says's that these variables are not even in policyengine-us!

6. 3 variables don't exist in policyengine-us

roth_ira_distributions, regular_ira_distributions, other_type_retirement_account_distributions are not in the TBS.
They'll always be zero-filled. The code handles this gracefully via the valid_outputs filter, but consider removing
them or adding a comment.

I know they've been "impossible targets" in calibration.

[baogorek]

CC:
All 4 SS sub-components are already handled by reconcile_ss_subcomponents() (puf_impute.py:376), which predicts
shares via QRF, normalizes them to sum to 1, and scales to match the PUF-imputed total. Stage 2's generic QRF would
overwrite with unconstrained values that likely don't sum to social_security.

Overlapping variables:

• social_security_retirement
• social_security_disability
• social_security_dependents
• social_security_survivors

[baogorek]

@MaxGhenis I 'm going to give you a preemptive check mark since I know you're working through the feedback and I've got to leave the house for a few hours.

Also, I'm eager to see the full publication process happen again.

[baogorek]

Merged main into this branch to catch up after PR #538 (calibration pipeline improvements). No conflicts — merged cleanly.