On the Onset of Robust Overfitting in Adversarial Training