Approximation and convergence of GANs training: an SDE approach