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 tidymodels/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!

[simonpcouch]

Ah, figured it out🙈 Wrapping in suppressMessages pending https://github.com/holub008/xrf/pull/21 merged + on CRAN.

[github-actions[bot]]

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.