Improving Variational Autoencoders with Density Gap-based Regularization