Intellect-2: A Reasoning Model Trained Through Globally Decentralized RL