Pytorch version (work in progress)

[florian-huber]

Several larger changes:

• New PyTorch implementation of the MS2DeepScore model
• New data generator for training the PyTorch model:
  • Uses the pre-selected spectrum pairs
  • Does the spectrum binning
  • Converts spectra to pytorch tensors
• Custom loss function combining MSE with bell curve to counteract predictions towards the mean plus a linear term to focus more on higher scores.
• Faster, numba based spectral pair selection