Rethinking Style and Content Disentanglement in Variational Autoencoders