DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning Hao Bai 1,2 Yifei Zhou 1 Jiayi Pan