Support for feature-wise transformations
Add the option to condition the output of the model using feature wise transformations, as described here. The current implementation only allows for the standard concatenation based approach. This MR gives the user greater freedom to experiment with various conditioning mechanisms, by giving the option to scale and/or bias the features of the input layer (prior to the initialisation network) and/or the output of the encoder layer, using the conditioning parameters.
Additionally this MR fixes some bugs in the parameterisation implementation that were introduced through previous updates of Salt.