Uncertainty-Aware Learning Against Label Noise on Imbalanced Datasets