Accept non-mlr3 fitted model

[jemus42]

Two possible routes:

1. Official mlr3 machinery to promote a fitted model to the corresponding mlr3 learner
2. Learner-specific predictor-function wrappers like in iml, mlr3summary etc.

If 1) is an option, it would definitely preferred, but this is unclear for now.

Before this happens we need to address #4, such that infrastructure is in place to use a trained learner and a dedicated test dataset (or still a task uses as full-on test-set?)

[jemus42]

Some learners like xgboost don't store their parameters such that they can be retrieved later, so from a fitted xgboost model it's probably not possibly to faithfully reconstruct the corresponding mlr3::Learner.

That wouldn't be too much of an issue though if we can at least get $train() and $predict() to work reliably, I guess?