LongRM: Revealing and Unlocking the Context Boundary of Reward Modeling