On the Statistical Capacity of Deep Generative Models