Tuning ML-pipelines without Learners

[mb706]

We might want to tune the behaviour of an ML pipeline without fitting a learner on it, e.g. a preprocessing pipeline that creates certain features where we want to optimize a property of the correlation matrix of the generated features. This might be something evaluated on the output of a pipeline (e.g. task generated by a preprocessing pipeline) or on the $state of the PipeOps involved, or both. It might not need to call a 'predict()', but it could still be something that is aggregated across CV folds, so would benefit from making use of the parallelization feature in mlr3tuning which parallelizes batch evals and CV folds at the same level. As I understand it, one would currently need to either write some kind of bogus-learner for this to use mlr3tuning, or write a custom function in bbotk::ObjectiveRFun[Dt].