Parallel Scaling Law for Language Models