Support for single fold run

[AlessandroLovo]

Allow to run a single fold instead of the standard k fold cross validation.
Also maybe return the trained model

see #59

[AlessandroLovo]

It is possible to run a single fold using the min_folds_before_pruning=1 and prune_threshold=-np.infty.
This ensures the run will be always pruned after only one fold.
Then you can use the nfolds and val_folds to tune your train val split. For example nfolds=10, val_folds=1 (default) uses 90% of the data for training and 10% for validation. More in general the validation fraction is val_folds/nfolds, so for example with nfolds=10, val_folds=3 you can achieve a 70-30 split.

This doesn't return the trained model though