Prune down what goes into the fit

[scarlehoff]

See here #2217 (comment)

This is due to the fact that by the time we go into the fit we still have access to all information from loading the datasets (the fktables in terms of huge pandas dataframes, loaded with and without cuts, same for the data, etc).
We don't need any of that during the fit, so we should drop all that before getting to the hyperparametrizable function.

This doesn't really make a difference in a normal fit (probably it is using more memory than needed but nothing too serious) but when doing a hyperparameter scan the memory starts becoming a bottleneck.

Not something urgent, but it would be a nice addition.