Distilling Reasoning Ability from Large Language Models with Adaptive Thinking