Improve early stopping of HGBT for scorers passed as string

As @NicolasHug mentioned in https://github.com/scikit-learn/scikit-learn/pull/18659#issuecomment-715522394, `HistGradientBoostingClassifier` and `HistGradientBoostingRegressor` need to recompute the predictions during training if one passes a string or callable to `scoring`.
At least for strings, we could implement a shortcut and use the underlying metric similar to how we use the loss.