Segmentation with mixed supervision: Confidence maximization helps knowledge distillation