[R] Add training metrics monitoring for xgboost()

[david-cortes]

ref #9810

I just realized that the new xgboost() interface was not doing anything with the verbose parameter. This PR adds an option to use the same x/y training data for metrics monitoring, which gets activated when passing a non-silent verbosity level.

[trivialfis]

Hmm, I missed this the last time. How does one use validation with this X/y interface?

[david-cortes]

There's no validation. This is just monitoring on the same training data.

I guess evaluation on a different dataset could be added later, but it would need be more involved - for example, it would need to re-encoded categorical columns in the evaluation data using the same categories as in 'x', would need to check whether the classes in 'y' match, and so on.

[trivialfis]

Thank you for the explanation. Might be necessary though, considering validation is pretty much the only effective way in keeping ML training in check. I doubt that anyone who uses ML training aside from DL doesn't have validation during training.

[david-cortes]

There's xgb.cv though. There could be an xgboost.cv (or cv.xgboost) to do it better, or somethng like 'eval_data_fraction' (better yet if implemented in the core library), but I think those could be added later after CRAN release blockers are addressed.