Use neural networks to predict the longitudinal hit position

Add a new class (TFCSPredictExtrapWeights) to predict the longitudinal hit position for each particle (photon, electron, charged pion) and calorimeter layer. This is done by using a neural network using as input the energy per layer, true particle's energy (and for photons and electrons, also the pdgId).

There is also some reorganization on TFCSEnergyAndHitGAN.cxx such that I can get energy per layer before simulating the hits.