Vectorization issue with ConditionalRealNVP

[sachaguer]

When evaluating the network using a RealNVP, vectorization does not work to perform the evaluation on multiple observation efficiently.

[sachaguer]

It is possible to circumvent the issue by creating a vmapped lambda function calling posterior.sample. It would be nice though to have an API that has the same properties no matter the neural network considered.