leaf Variable inplace bug fix

[glenn-jocher]

This PR fixes Docker issue with leaf Variable inplace ops in ultralytics/yolov5#1552.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improvement in bias initialization process within the YOLOv3 model.

📊 Key Changes

• Adjustments made to the way biases are initialized in the Detect() module.
• Use of b.data instead of b when modifying bias values.

🎯 Purpose & Impact

• Enhanced Stability: Direct modification of b.data ensures that gradient tracking is not affected during bias initialization, leading to more stable training behavior.
• Potential for Better Performance: Proper initialization of biases can contribute to faster convergence and potentially better object detection performance for users.
• Backward Compatibility: These changes maintain compatibility with existing model configurations and training processes, enabling users to benefit without altering their existing workflows.