train_loader lazy

[lazarusA]

currently computing losses for validation and training (fully) arrays must be fully in memory, we should change this to allow for lazy evaluation, in batches.