Coresets for Robust Training of Neural Networks against Noisy Labels