Deconstructing In-Context Learning: Understanding Prompts via Corruption