Variational autoencoders in the presence of low-dimensional data: landscape and implicit bias