Fix ONNX --dynamic export on GPU

[glenn-jocher]

Patch forces --dynamic export model and image to CPU. Resolves bug raised in #8377

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced ONNX export compatibility and adjusted half precision handling during export.

📊 Key Changes

• Modified the export_onnx function in export.py to move models and input tensors to CPU when using the --dynamic flag.
• Changed the order of operations in the run function to apply half precision (FP16) to the model and input only after initial model dry runs.

🎯 Purpose & Impact

• The update to the ONNX export function ensures that dynamic exports work as intended, as dynamic operations are only compatible with CPU. This allows for smoother conversion of models to ONNX format when using dynamic axes. 🔄
• Adjusting when half precision is applied prevents potential issues during model export, such as incompatibilities with certain export targets like CoreML and XML which do not support FP16. This ensures more stable and reliable model export, particularly beneficial for users working with mixed precision to optimize performance. 🏃‍♂️💨