Are More LLM Calls All You Need? Towards Scaling Laws of Compound Inference Systems