Describe the issue linked to the documentation

This issue concerns:

• sklearn.ensemble.RandomForestClassifier
• sklearn.ensemble.RandomForestRegressor
  and probably:
• sklearn.ensemble.ExtraTreesClassifier
• sklearn.ensemble.ExtraTreesRegressor
• sklearn.ensemble.BaggingClassifier
• sklearn.ensemble.BaggingRegressor

The documentation of oob_decision_function_ in RandomForestClassifier is:

Decision function computed with out-of-bag estimate on the training set. If n_estimators is small it might be possible that a data point was never left out during the bootstrap. In this case, oob_decision_function_ might contain NaN. This attribute exists only when oob_score is True.

The description of oob_prediction_ RandomForestRegressor is:

Prediction computed with out-of-bag estimate on the training set. This attribute exists only when oob_score is True.

However, PR #19162 changed the computation of OOB scores and NaN have been replaced with 0. The description in RandomForestRegressor does not even mention the issue that a data point could never be left out.

The PR also dropped support for multiclass-multioutput so I'm not sure if the shape of oob_decision_function_ may still be (n_samples, n_classes, n_outputs) (not 100% sure on this one).

Suggest a potential alternative/fix

For random forest estimators, the new descriptions could be something like this:

oob_decision_function_ : ndarray of shape (n_samples, n_classes)
        Decision function computed with out-of-bag estimate on the training
        set. If ``n_estimators`` is small it might be possible that a data point
        was never left out during the bootstrap. In this case, the probabilities
        for such data points to belong to each class are 0. This attribute exists
        only when ``oob_score`` is True.

and

oob_prediction_ : ndarray of shape (n_samples,) or (n_samples, n_outputs)
       Prediction computed with out-of-bag estimate on the training set. If 
       ``n_estimators`` is small it might be possible that a data point was never
       left out during the bootstrap. In this case, the predictions for such data points
       are 0. This attribute exists only when ``oob_score`` is True.

The descriptions for the other estimators may also require some updates.