Would also require the user to supply test data, e.g. as a new task or presumably just as a data.frame.
Would then require conversion to a Task under the hood and manually ensuring everything stays consistent (column names/types/order).

Needs distinction for PFI/CFI/RFI and SAGE, where we don't need to retrain the learner and would just need to decide how to handle resamplings and LOCO/LOCI where refitting is kind of the point.

No refitting needed

• Accept trained learner + task + instantiated resampling
  • -> predict / evaluate on the test-sets?
• Accept trained learner + test_dt data.frame, no resampling
  • -> predict /evaluate once on dedicated test dt via $predict_newdata but would still need a Task object, which we need to create on the fly or assemble with additionally provided information (target column, positive class for bin classif, etc.)

Refitting needed

• Accept trained learner + task / resampling (or not, see above)
  • Use trained learner as reference model for LOCO, reset and retrain for feature subsets as currently done
  • Not really possible for LOCI I guess, so we don't need to bother?