consider moving preprocessing "into model"

[sammlapp]

With Tensorflow, preprocessing actions can be added into the model and run on GPU - can we do this in Pytorch? How much flexibility would we sacrifice with preprocessing options? Would it allow us to export models in a shareable cross-platform way?

[sammlapp]

Initial experiments show that tensor operations could happen on GPU within the pytorch module and that this would make them very fast.

[sammlapp]

Additionally, having preprocessing (in particular FFT) inside the PyTorch model object will help when we want to make the models shareable between platforms/packages (eg with ONNX)

[sammlapp]

see #556