Replaying pre-training data improves fine-tuning