Command injection and Path traversal security fixes

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced security and reliability in distributed training setup.

📊 Key Changes

• Hashing function switched from MD5 to SHA256 for file integrity checking.
• Distributed training command generation now avoids security vulnerabilities.
• Temporary files created during distributed training are cleaned up more reliably.
• Introduced compatibility with torch.distributed.run for PyTorch 1.9.

🎯 Purpose & Impact

• 🔒 Increased Security: By upgrading to SHA256, the codebase benefits from a more secure method of hashing, reducing vulnerability to hash collision attacks.
• 🛠️ Improved Reliability: Generating a proper file for distributed training and cleaning up temporary files prevent potential issues that could arise from improper handling of resources during the training process.
• 🤝 Compatibility and Flexibility: Ensuring support for PyTorch's torch.distributed.run method provides better compatibility with newer versions of PyTorch, future-proofing the project.
• 🚀 Potential Impact: Users of Ultralytics can expect a more robust training environment, especially when training models in distributed settings across multiple machines or GPUs.