Capabilities and Limitations of Time-lagged Autoencoders for Slow Mode Discovery in Dynamical Systems