Hub device mismatch bug fix

[glenn-jocher]

This PR attempts to fix issue #1617.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improvement to non-maximum suppression function in YOLOv5's utility code.

📊 Key Changes

• Modified the way the number of classes (nc) is determined from the predictions to ensure compatibility.
• Optimized the initialization of the output tensor to leverage the device of the input predictions, improving performance.

🎯 Purpose & Impact

• 🚀 The fix ensures that the number of classes is correctly calculated for any given input shape, which enhances the model's flexibility and robustness.
• 🌐 By initializing the output tensor on the same device as the input, it eliminates unnecessary data transfers between devices (e.g., CPU to GPU), leading to faster and more efficient inference.
• ✅ These changes will potentially benefit all users by providing more accurate object detection and improving the speed of the YOLOv5 model inference, without altering the model's core functionality.