FLEURS: FEW-Shot Learning Evaluation of Universal Representations of Speech