Issue with `KerasModelProblem` and variational autoencoders

[tomoleary]

Need to work through the kinks of the construction of the KerasModelProblem class when the neural network has complex callable structure. This could be a problem for generative samplers in general.

In the meantime the VariationalAutoencoderProblem class works.