Preserving Generalization of Language models in Few-shot Continual Relation Extraction