FeaturelessLearner: if "learner_weights" is added, the FFL should support weights too

[berndbischl]

otherwise we cannot use it nicely as a fallback

for some of the "weirder" options of the FLL this might be harder, but at least for this it should be doable:

regr: mean prediction
classif: mode prediction, and prob estimation.

i would also vote to remove some "weirder" options from the learner to make it simpler

[berndbischl]

should be obvious: if this is added, add unit tests