Understanding Variational Autoencoders with Intrinsic Dimension and Information Imbalance