Prefer MPS over CPU if available

[glenn-jocher]

Attempts to prefer Apple MPS if supported and available over CPU for all YOLOv5 ops. May resolve #8207

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improvements in device selection logic for YOLOv5 model loading.

📊 Key Changes

• Refactored device selection code in hubconf.py by removing a redundant check that defaulted to CUDA if available.
• Updated select_device() function in utils/torch_utils.py:
  • Simplified the CUDA device selection condition.
  • Included a check for Apple's Metal Performance Shaders (MPS), leveraging GPU-like acceleration if available and not on CPU.
  • Automatically defaults to using the CPU if neither CUDA nor MPS is available.
  • Cleaned up code readability and documented preference order (CUDA > MPS > CPU).

🎯 Purpose & Impact

• 🎨 Streamlines the device selection process for users, offering a clearer logic flow.
• 🏃‍♂️ Speeds up initial loading of the YOLOv5 model by prioritizing GPU usage when available for better performance.
• 🍏 Exploits MPS on compatible Apple devices, enhancing performance on Mac hardware where MPS is supported.
• 🆕 Ensures compatibility across a wider range of devices by properly handling device options and fallbacks.
• 💻 Users without GPU support will automatically fall back to CPU without needing extra configuration.