Reminding the Incremental Language Model via Data-Free Self-Distillation