Enhancing Auto-regressive Chain-of-Thought through Loop-Aligned Reasoning