Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage Policy Optimization