GRACE: Generating Socially Appropriate Robot Actions Leveraging LLMs and Human Explanations