Beyond temperature scaling: Obtaining well-calibrated multiclass probabilities with Dirichlet calibration