Updates for MT classification model performance evaluation

[paulsonak]

MT classification models have two problems:

1. the weights aren't calculated or created in MinimalDataset used for calculating metrics on the original training data. So, the metrics and predictions fail
2. Compare models doesn't let you pull multiple single-task model metrics from models trained on one multi-task dataset with a single response column passed. This is useful for using a MultitaskScaffoldSplit on the dataset and just general ease when training models on dense multitask datasets.

[paulsonak]

Updated PR to integrate into bug_transformer_fitting: #387

[paulsonak]

Additional problem: the get_multitask_perf_from_files_new() function could not handle a mix of single and multitask model metrics in the same directory search.