xgboost and categorical predictors

There is an experimental argument for [categorical variables](https://xgboost.readthedocs.io/en/stable/tutorials/categorical.html). 

We need a process to potentially adopt this and how we should handle

* backward compatability
* changes to `parsnip:::as_xgb_data()` since some new xgboost options are required for this. 
