Topological Obstructions to Autoencoding