Consistent handling of missing cross-validation class information

Develop a consistent treatment of missing class information across the Ridge, Extra Trees, and Naive Bayes base learners in TIGER when cross-validation folds omit a class from their training data, rather than discarding Ridge scores and silently zero-filling the corresponding Extra Trees and Naive Bayes scores.

Background

For data sets containing a class with only one training instance, TIGER cannot use stratified cross-validation to ensure that every fold contains every class in its training data. Consequently, some folds omit a class entirely. The current implementation handles this inconsistency differently across base learners: Ridge out-of-fold scores for affected cells are discarded, whereas Extra Trees and Naive Bayes scores are silently set to zero. This reduces TIGER from twelve to eight base learners on three benchmark data sets and leaves unresolved how the missing information should be represented or recovered uniformly across all learners.