OmniRL: In-Context Reinforcement Learning by Large-Scale Meta-Training in Randomized Worlds