Hub device mismatch bug fix

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved device handling in non_max_suppression function.

📊 Key Changes

• Changed the reference to the number of classes nc for consistency.
• Specified device argument when creating zero tensors, ensuring they're created on the same device as prediction.

🎯 Purpose & Impact

• Ensures that the class count (nc) is always accurately derived from the input predictions' shape.
• The update ensures that tensors created during non_max_suppression are allocated on the same hardware (CPU or GPU) as the input, leading to performance improvements and avoiding unnecessary device transfers which can be a bottleneck.
• This impacts users by potentially speeding up the non_max_suppression step of object detection, resulting in faster inferencing times and smoother integration with hardware accelerators like GPUs. 🚀