Per-model calibration

Apply calibration to each model in the ensemble to potentially improve results more than simply calibrating the final weighted ensemble. Has potential to improve metrics such as ROC-AUC, which is currently not improved via final weighted ensemble calibration.

Steps:
1. Try per-model calibration on non-bagged setting -> Temperature scale each model after it trains
2. Try to calibrate a bagged ensemble -> Either just use out-of-fold predictions for calibration or calibrate each child individually.
3. Try to calibrate end-to-end multi-layer stack ensemble -> Calibrate each bagged ensemble of the stack + weighted ensemble.

No files, tests, or entry points are named. Start by locating the model-training and ensemble-calibration paths, then compare the three proposed experiments: non-bagged temperature scaling, bagged calibration, and multi-layer stack calibration. Done means evaluating whether per-model calibration improves metrics such as ROC-AUC over final weighted-ensemble calibration.

Written by the indexing model from the issue text.