Add automatic confidence-threshold calibration to both training pipelines

[nilshof01]

…ines

The optimal mask threshold is not a constant: measured across 3 regions x 2 models x 5 folds (size 32) it ranged 0.10-0.90, and the shipped operating points (UNet fixed 0.5, DINOv3 catalog 0.4371) cost +1.1 to +13.9 test F1 vs a deterministic 17-point val-split sweep, which landed within 0.7 F1 of the test-oracle in every measured cell.

unet_segmentation:

• new calibrate_threshold step: val-split F1 sweep, rate-matching on the train sampler when val has no positive pixels (threshold at which the predicted positive-pixel fraction matches the labeled fraction)
• predict() now thresholds the most likely foreground class directly; the previous argmax+top-prob rule made binary thresholds below 0.5 no-ops (identical output at the 0.5 default)

dinov3s_buildings:

• new calibrate_threshold step (same sweep on the spatial val chips via cache_val_forwards; rate-matching on train chips when val < 2 chips)
• tune_postprocess defaults are seeded with the calibrated threshold, so the skipped-search path (val < 8 chips or trials disabled - i.e. every small-dataset run) serves the calibrated value instead of the catalog constant. When Optuna runs it still tunes the threshold jointly in [0.3, 0.8]; fixing it there would require a dinov3_hot change.

Both STAC items declare training.calibrate_threshold (default true).