Bugfixes

[LonnekeScheffer]

Fixes related to positive class not being set properly:

• Label: ensure label classes always follow the same order (positive class last, matches sklearn), utility function for retrieving negative class in binary case, enforce label classes/values to always be set
• LabelConfig: ensure a default label positive class is always set
• predictions proba: explicitly keep track of the label classes (dict instead of multi-dimensional np array) -> previously the ordering of labels (positive class last) was not consistently enforced across MLMethods. These bugs are dificult to catch as the 'predictions' were correct, but the 'predictions_proba' not. I believe the current solution is less error-prone in the future when other developers may work on immuneML.
• These updates resolve previously observed bugs that: predictions & predictions_proba did not match (resulting in inverse ROC curves), and that the wrong positive class was sometimes assumed for asymmetric performance metrics (e.g., precision, recall)

Other (minor) fixes:

• Fix in matches report: correctly retrieving subject id
• Bugfix ManualSplitter._make_subset: didn't work for non-string example ids
• ManualSplitter log text: shortened an excessively long log message
• MotifSeedRecovery: small update error message

[LonnekeScheffer]

Fixed one additional bug in MLApplication: positive_class was not tracked when exporting the ML settings. Tested the fix with real data and it works. This makes exported ML models from previous versions of immuneML technically incompatible when imported with the version after this bugfix, but it is too small for a 'major' version bump. I think very few people use MLApplication so far, so I think it should be fine with a minor version bump (2.3.0) or even patch (2.2.3).