Tune temporal batching for model optimization

[JordiBolibar]

This could be a key element in model training, since it acts as a hyperparameter.

Basically, we have two main strategies:

• Whenever we integrate and hit a seasonal period, we can add a new tree or backpropagate. And whenever we hit an annual period, we do the same. This means, to optimize the model for the smallest temporal period possible.
• Alternatively, we can do this in conjuction, and only optimize whenever we have integrated both the winter and summer seasonal data + the annual value. We compute a combined loss of these 3 values before optimizing the model. This introduces a new hyperparameter, which we should add the to the grid search.