Fix calibrator pickle compatibility and add clean inference API

[michaelzenkay]

Summary

• Fix CalibratedClassifierCV pickle loading across sklearn versions (0.20 → 1.7+) by patching renamed internal attributes (calibrators → calibrators_, classes → classes_, base_estimator → estimator)
• Add scripts/infer.py — a clean programmatic API (predict()) that builds args internally instead of requiring sys.argv manipulation
• Fix torch.load for PyTorch 2.6+ (weights_only=False)

Details

Calibrator compatibility (scripts/main.py)

Calibrator pickles saved with older sklearn versions fail on newer versions due to renamed internals:

Pickle saved with	Breaks on	Error
sklearn ~0.20	>= 0.22	ModuleNotFoundError: sklearn.svm.classes
sklearn 0.23	0.24+	_CalibratedClassifier missing calibrators attr
sklearn 0.23	1.2+	CalibratedClassifierCV missing estimator attr

Added _patch_calibrator() to fix attribute names after loading, and load_calibrator() with graceful error handling.

Clean inference API (scripts/infer.py)

New programmatic interface that avoids sys.argv manipulation:

from scripts.infer import predict

df = predict(
    metadata_csv="path/to/metadata.csv",
    model_dir="path/to/models/",
    calibrate=True,
)
# df has columns: patient_exam_id, 1_year_risk, ..., 5_year_risk

PyTorch 2.6+ compatibility (onconet/models/factory.py)

Added weights_only=False to torch.load calls since PyTorch 2.6 changed the default to True.

Test plan

• Load calibrator pickle on sklearn 0.24.2 and 1.7.0, verify predict_proba works
• Run infer.predict() on single exam, verify calibrated output
• Run scripts/main.py end-to-end, verify CSV export with calibrated scores
• Batch inference on 1k+ exams (Windows + Linux HPC), calibrated output verified