On Deep Generative Models for Approximation and Estimation of Distributions on Manifolds