Draft: probNN run3 v1.0 tuples + training

-- 1/05 --

Update on samples for training and also the training phase by itself.

Samples for training probNN:

• Electrons: B->JpsiK, ee
• Muons: B->JpsiK, mm + inclusive Jpsi, mm
• Pions: Ks->pipi (low momentum) + Bu->Kpi (high momentum) + few extraction from Lc and D* to enlarge samples
• Kaons: D*->D0pi, Kpi (all range) + D*->D0pi, KK (to enlarge sample)
• Protons: Lc->pKpi, Lb->ppi

Regarding the training, a learning rate scheduler to help converging along with an early stopper to prevent overfitting have been included. All together helps the convergence of the models, pushing them 2-3% further with fewer iterations (generally a third of epochs). This is parametrized by hand, we could be as restrictive as desired but the current configuration provides quite a good balance of allowing model explore more regions before stopping, and stop too early.