Local Convergence of Gradient Descent-Ascent for Training Generative Adversarial Networks