Add option to disable drop last on training

[marco-garosi]

Hello, I appreciate your toolbox and I am actively using it! I have implemented a small change in the data loader construction in data_manager.py to force disabling drop_last on training datasets from the configuration file via the new DISABLE_DROP_LAST option. In some cases, one may not want to drop the last training batch, so this option allows the user to choose so. Note that the original behavior is left untouched by default, as this additional option has to be activated explicitly -- thus, it won't break existing code. I have implemented it for TRAIN_X, but it may be implemented for TRAIN_U as well.