leaf Variable inplace bug fix

[glenn-jocher]

This PR fixes Docker issue with leaf Variable inplace ops in #1552.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved bias initialization in YOLO model.

📊 Key Changes

• Modified the code to use b.data instead of b when adjusting bias values during model initialization.

🎯 Purpose & Impact

• 🎯 Purpose: To strictly modify the tensor data without affecting its grad_fn, thus enhancing the code's consistency and clarity when initializing biases.
• 🎈 Impact: This tweak ensures that gradient calculations are not inadvertently affected during initialization, potentially leading to more stable and reliable training of the YOLO models for object detection tasks. Users may experience subtle improvements in model training efficiency and performance.