On class-imbalanced datasets, Macro F1 is extremely sensitive to rare class performance, where small threshold changes cause dramatic fluctuations in precision and recall for rare classes