Forward Super-Resolution: How Can GANs Learn Hierarchical Generative Models for Real-World Distributions