[air] Improvements to XGBoost Benchmarking script

[richardliaw]

• Could we have separate scripts for training and prediction? This is for readability: it's not great to have multiple workflows in the same script.
• Should make @run_and_time_it a common util
• to_air_checkpoint(".", model) <--- isn't the "." fishy?
• batch_predictor.predict(ds.drop_columns(["labels"])) <--- I think you can use the "feature_columns" arg of predict() that Amog added, instead of dropping labels.

Originally posted by @ericl in #26608 (review)