TensorRT PyTorch Hub inference fix

[glenn-jocher]

Solution proposed in #7128 to TRT PyTorch Hub CUDA illegal memory errors.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improvement in device compatibility for AMP inferences within YOLOv5.

📊 Key Changes

• Ensured that the dummy tensor p is on the same device as self.model when self.pt is False.

🎯 Purpose & Impact

• Purpose: To fix a potential issue where the dummy tensor for device and type determination was not being consistently placed on the correct device (e.g., GPU or CPU).
• Impact: Enhanced compatibility and stability when running Automatic Mixed Precision (AMP) inferences, preventing errors related to device mismatches. This change is particularly relevant when self.pt is False, ensuring that users who are performing inferences without a .pt model (PyTorch format) do not face device-related issues.

[glenn-jocher]

Fix works!