"Well, Keep Thinking": Enhancing LLM Reasoning with Adaptive Injection Decoding