Autoencoding Dynamics: Topological Limitations and Capabilities