Using custom latent sample for GAN training

[mbarbetti]

As discussed with @cpviolation, deep generative models (e.g., GANs, Normalizing Flows) can be used to parameterize the transformation from a probability distribution to another, once instances sampled from both the distributions are available in the training sample. To unlock this nice property, PIDGAN should provide a new class of generators that can take directly as input elements from the latent space. Such latent space represents a proxy for the source probability distribution and the generator should be rewritten to describe a transformation for the target probability distribution.