Fast Adversarial Training with Smooth Convergence