Feature Request: Allow no required predict sets for a `Measure`

[sebffischer]

Some mlr3 Measures don't require a prediction object to operate on, such as msr("selected_features") or msr("internal_valid_score").

Let's say I want to resample xgboost with early stopping and I only care about the internal validation scores (e.g. in the context of tuning). I could then avoid making predictions on the test set alltogether. What I want to do is therefore:

l = lrn("classif.xgboost", nrounds = 100,
  early_stopping_rounds = 10, validate = "test", predict_sets = NULL)
  
rr = resample(tsk("iris"), l, rsmp("holdout")

rr$aggregate(msr("internal_valid_score"))

The reason why the code above does not run is because each measure must have at least one required predict sets because of this check:

mlr3/R/Measure.R

Line 147 in 3f159d9

self$predict_sets = assert_subset(predict_sets, mlr_reflections$predict_sets, empty.ok = FALSE)

I.e., for MeasureInternalValidScore I need to pick "test", "train" or "internal_valid", even though none are required.
The last call to $aggregate() will therefore err, no matter how I define the MeasureInternalValidScore.

[mllg]

You can toggle empty.ok and check if the unit tests still pass. But to be fair, not wanting to look at predictions at all is an unusual setting ;)