Making Language Models Better Reasoners with Step-Aware Verifier