Improving Factual Error Correction by Learning to Inject Factual Errors