Dynamic Chain-of-Thought: Towards Adaptive Deep Reasoning