Improving Channel Charting with Representation-Constrained Autoencoders