Analyzing and Improving Adversarial Training for Generative Modeling