On the capacity of deep generative networks for approximating distributions