Stepwise Informativeness Search for Improving LLM Reasoning