importance filter default is bad

[mb706]

the importance filter by default uses the "featureless" learner by default, which makes the filter basically useless. There are likely very few cases in which the user wants to initialize the importance filter with the featureless learner. Therefore, the learner should not have a default; when constructing the importance filter without providing a learner, an error should be thrown instead.

[sebffischer]

I agree: https://github.com/mlr-org/mlr3filters/issues/130

[sebffischer]

I think these choices were made to make the as.data.table() conversion work. Maybe we should implement your solution from mlr3pipelines in mlr3misc.

[sebffischer]

https://github.com/mlr-org/mlr3misc/pull/92 would help for the dict -> dt conversion