Better Uncertainty Calibration via Proper Scores for Classification and Beyond