Update on the NN fitting

• Tried directional fitting of momentum vector: pX/pZ & pY/pZ
  • Did not work as there can be tracks with p/pT ~1  -> pZ is basically 0 -> ratios explode
• Changed fitting to pX/p, pY/p, pZ/p
  • Improved training quality and convergence behaviour significantly
  • Regained the position resolution comparable to native clusterizer as before with additional momentum vector estimation

• Momentum vector estimation works (somehow I messed up the training in Y and Z though, so need to redo it)

Otherwise busy writing some clusterization kernel code and further sw testing for network implementation with GPU backend