Change the hyperoptimization target minimization from runcard

Created by: scarlehoff

The options for now are average, std and best_worst as those are the ones included in the documentation.

On doing this I found a very important bug (hyperopt only), the positivity and integrability multiplier was not being reset for each fold so effectively by the final fold the fit needs to be strictly positive and integrable as it has an almost infinity multiplier @RoyStegeman I think this is the reason you saw the problems you saw. Sorry.

A funny byproduct has been that the preferred fits were those finishing early: large number of epochs but very little patience. I guess because that way the "infinite multiplier" doesn't happen.