Exploring the Robustness of Decentralized Training for Large Language Models