Characterizing and Avoiding Problematic Global Optima of Variational Autoencoders