TensorBoard DP/DDP graph fix

[glenn-jocher]

Possible fix for #3284.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced model compatibility in train.py for single-GPU environments.

📊 Key Changes

• Replaced is_parallel() with a new function de_parallel() in train.py and torch_utils.py.
• Updated usage in train.py to de-parallelize the model for certain operations.

🎯 Purpose & Impact

• Simplifies model handling: Ensures a consistent single-GPU model is used for operations like TensorBoard graph addition and model checkpoints, preventing errors in non-parallel environments. 🖥️
• Better code clarity: The de_parallel function clearly communicates its purpose and makes the codebase more maintainable. ✨
• User experience: Streamlining the code can help users avoid common pitfalls when moving between parallel and non-parallel training setups. Uses of the model in different parts of the code are less error-prone. 👥