On the Convergence and Robustness of Training GANs with Regularized Optimal Transport