prefit xgboost model in xrf_fit(), allow early stopping

[simonpcouch]

This PR:

• Adds support for early stopping with xrf::xrf()
• Corrects routing of mtry from colsample_bytree to colsample_bynode (and is now thus consistent with parsnip, see mtry maps to wrong parameter for XGBoost parsnip#495)
• De-duplicates xgboost training code between parsnip and rules
• Extends testing for xrf engine and removes mtry tests that are now redundant with parsnip

It does so by making use of the prefit_xgb argument to xrf::xrf()—instead of having xrf handle the xgboost training, we use parsnip::xgb_train and pass the pre-fit xgboost model to xrf.

Justification for pre-fitting

The pre-fitting makes some parts of wrapping easier for us, but was initially a need for early stopping:

xrf::xrf.formula() wraps both xgboost::xgb.train() and glmnet::glmnet(). xrf::xrf.formula() takes in an xgb_control argument, where all arguments passed to xgb_control are passed to xgboost::xbg.train()'s param argument—this is currently the only way one can pass arguments to xgboost::xgb.train() through xrf::xrf.formula().

This is an issue for early stopping, as early_stopping_rounds is a main argument to xgboost::xgb.train() and cannot be passed through param. Thus, if we wanted to implement an interface for early stopping, we'd either:

• need to contribute to the xrf package to allow for passing arguments to xrf::xrf.formula() as main arguments to xgboost::xbg.train()
• "prefit" the xgboost model with our own machinery and pass it as prefit_xgb to xrf::xrf.formula().

Related to tidymodels/parsnip#749!

[EmilHvitfeldt]

Couple of small changes to the tests and I think we are good here!

[simonpcouch]

Ah, figured it out🙈 Wrapping in suppressMessages pending holub008/xrf#21 merged + on CRAN.

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.