better thread efficiency for cv prunning with thresholds

We could have better thread efficiency for early stopping with thresholds (when not used with selection early stopping). Currently, all threads need to come together at the end of each round of CV. Would be better if a single thread would evaluate a all the folds until termination (the same as regular evaluation).

related to #69 . It may be better to implement this as a standalone function or create a pruning API.