This is finally supported in mlr3 and mlr3learners for xgboost (currently using the latest github versions).

See example here. Setting offset in the task means also that it is applied automatically during internal tuning/validation of xgboost as well.