This PR considers a helper for multilabel/set-wise evaluation metrics such as precision, recall, fbeta, jaccard (#10083), fall-out, miss rate and specificity (#5516). It also incorporates suggestions from #8126 regarding efficiency of multilabel true positives calculation (but does not optimise for micro-average, perhaps unfortunately). Unlike confusion_matrix it is optimised for the multilabel case, but also handles multiclass problems like they are handled in precision_recall_fscore_support: as binarised OvR problems.

It benefits us by simplifying the precision_recall_fscore_support and future jaccard implementations greatly, and allows for further refactors between them. It also benefits us by making a clear calculation of sufficient statistics (although perhaps more statistics than necessary) from which standard metrics are a simple calculation: it makes the code less mystifying. In that sense, this is mostly a cosmetic change, but it provides users with the ability to easily generalise the P/R/F/S implementation to related metrics.

TODO:

• implement multilabel_confusion_matrix and use it in precision_recall_fscore_support as an indirect form of testing
• fix up edge cases that fail tests
• benchmark multiclass implementation against incumbent P/R/F/S
• benchmark multilabel implementation with benchmarks/bench_multilabel_metrics.py extended to consider non-micro averaging, sample_weight and perhaps other cases
• directly test multilabel_confusion_matrix
• document under model_evaluation.rst
• document how to calculate fall-out, miss-rate, sensitivity, specificity from multilabel_confusion_matrix
• refactor jaccard similarity implementation once [MRG] average parameter for jaccard_similarity_score #10083 is merged

If another contributor would like to take this on, I would welcome it. I have marked this as Easy because the code and technical knowledge involved is not hard, but it will take a bit of work, and clarity of understanding.