A Distilled Representation for Zero and Few-Shot Localization of Task-Oriented Dialogue Agents