Detection() device bug fix

[glenn-jocher]

This PR fixes a CUDA device bug with PyTorch autoshape models with Detection() outputs.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhancement in tensor device handling for bounding box normalization.

📊 Key Changes

• Device information is now explicitly assigned to normalization gains tensors.
• torch.Tensor replaced with torch.tensor to leverage the device argument.

🎯 Purpose & Impact

• 🛠️ Purpose: To ensure that tensors used for bounding box normalization are created on the same device as the predictions to avoid device mismatch errors.
• 💻 Impact: Provides a more reliable and consistent tensor handling process, particularly when working with models across different devices. This can lead to smoother model deployments and less troubleshooting for developers and users.
• ⚡️ User Experience: Improved performance when processing images across CPU and GPU, with reduced risk of encountering device-related bugs.