Enable weighted sampling of features

[lutzvdb]

In our random forest models (using the R package "ranger") we frequently use weighted sampling of features (argument "split.select.weights", compare official documentation here ). I would like to apply the same logic to xgboost trees, as in my usecase I frequently know which variables play a more important role for a certain forecast period (we do time series analysis). I already utilize feature sampling using colsample_bytree and colsample_bynode, so I would greatly appreciate being able to alter the probabilities for column sampling.

[hcho3]

Related: #3754, #4230

[trivialfis]

I will look into this.