Jump Starting Bandits with LLM-Generated Prior Knowledge