Purify Unlearnable Examples via Rate-Constrained Variational Autoencoders