Early stopping when wrapping an `XGBClassifier` into a `OneVsRest` meta-estimator

Question

I am trying, without success, to use early stopping on an XGBoost classifier applied to a multilabel problem using the Scikit-Learn API. To provide an example, the snippet b