Towards a Theoretical Understanding of the Robustness of Variational Autoencoders