Prototypical Calibration for Few-shot Learning of Language Models