Benign Overfitting in Classification: Provably Counter Label Noise with Larger Models