ComputerRL: Scaling Reinforcement Learning for Computer Use Agents