The error received is indicative of the issue. For trees of the class ClassificationPartitionedEnsemble, you cannot use prediction, so this kind of tree does not have a predict method and hence, cannot be used with 'predictorImportance'.
fitcensemble returns ClassificationPartitionEnsemble object as shown below.
Use kfoldPredict, kfoldLoss, kfoldMargin, kfoldEdge, and kfoldfun in order to use estimate the quality of the classification using fitcensemble.
