Fine-tuned Language Models are Continual Learners