Does Thinking More always Help? Understanding Test-Time Scaling in Reasoning Models