Stability and Generalization of Adversarial Training for Shallow Neural Networks with Smooth Activation