Compression of Structured Data with Autoencoders: Provable Benefit of Nonlinearities and Depth