selected_features for learners that don't support it should be the entirety of features seen in training

[mb706]

This way we could correctly query a pipeline that selects features first and gives the result to a learner. The GraphLearner could then ask the learner at the end how many features it used, and if it is a learner that supports embedded featsel (rpart e.g.) then this would give the correct value, but even for learners that do not do support it the result could make sense.

Also this would solve https://github.com/mlr-org/mlr3fselect/issues/87