Control via YAML files

[SCiarella]

It can be useful to collect all the relevant variables like:

• gridsize
• dt
• Nsimulations
• loss subparameters (nunroll, ecc..)
• NN model for closure

into a single configuration_x.yaml. This will allow us to run generate.jl configuration_x.yaml or postprocess.jl configuration_y.yaml on a server without using a REPL. This approach is better than training all the different models via a single train_array.jl because it allows more flexibility.

[luisaforozco]

I see that the current code is prone to error in the following case: train a-priori, the batch_size is used (i.e. manually and explicitly specified) in preprocess_priori.jl to create the dataloader. Then in train_priori.jl for the creation of the callback there is a new batch_size that needs to be passed. I do not think is problematic if they do not match exactly but is weird that we do not have it as a global param of the training.