Contrastive Learning Improves Model Robustness Under Label Noise