Pass Errors to the tuner

Sometimes there are some hyperparameter combinations that provoke errors, so we know that those are bad values, but we don't have any score to pass to the tuner.

Should we implement a way to pass these errors to the tuners that allows them to figure out which hyperparameter values provoke them?

This is a possible implementation:

• Allow passing None as the score, which means that score could not be computed for those hyperparameter values.
• Filter out all training samples with score None when fitting the current estimators.
• Implement a new binary classifier that sits before the current estimators and only predicts whether the score will be None or not, and use the predictions from this classifier to discard suggested hyperparameters during propose.