ensure correct prediction types are used when postprocessing

[topepo]

tune currently uses the specified metric sets to determine what types of predictions are being made. For example, for a binary classification model, we might only request "class metrics" (e.g. sensitivity, accuracy, etc), tune will not generate probability predictions.

However, some postprocessors require multiple types of predictions. For the above example, we might want to optimize the probability threshold. Even though we are only modifying the class predictions, we need the probability predictions to do so.

So: tune should use the metric set and any postprocessor to determine the required prediction types.

[topepo]

Note that the adjustment objects have elements that list the type of prediction is required

[topepo]

Also, we'll need to take into account the outputs too. For example, calibration uses probability predictions to adjust all classification prediction types. If we do not have hard class prediction columns, and error will occur (although we could change this in probably)