Generalization Bounds of Stochastic Gradient Descent for Wide and Deep Neural Networks