Cache QRF models to avoid retraining when only weights change

[MaxGhenis]

Summary

The extended CPS build retrains all QRF models from scratch every time, even when only calibration weights change. Since the QRF imputation depends on source CPS + PUF data (not weights), the fitted models could be cached and reused.

Current cost

• 85+ variable sequential QRF on ~20K PUF records: ~30-60 min
• Additional QRF calls for weeks_unemployed, retirement contributions, SS sub-components
• This runs on every make data or CI build

Proposed approach

• Serialize fitted QRF models (e.g. pickle/joblib) keyed by a hash of the training data
• On rebuild, check if source data hash matches cached model — if so, skip training and just predict
• microimpute could potentially support this natively (save/load fitted models)
• Could also cache the full extended_cps_2024.h5 and only rebuild when CPS/PUF inputs change

Context

Related to the sequential QRF migration in #594 — now that all 85 variables are in a single fit() call, caching the one fitted model would skip the entire training phase.