Add macos-latest runner for CoreML benchmarks

[glenn-jocher]

Signed-off-by: Glenn Jocher glenn.jocher@ultralytics.com

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced model output handling for segmentation in YOLOv5.

📊 Key Changes

• Simplified the logic for handling outputs in models that include segmentation tasks.
• The previous method of identifying the last layer's output by key has been replaced.
• Outputs are now retrieved by reversing the order of model values which is particularly relevant for certain segmentation models.

🎯 Purpose & Impact

• The change aims to streamline the post-processing step in YOLOv5's forward pass, especially for models that perform segmentation tasks.
• This can potentially make the code easier to maintain and understand, as it removes the need for dynamically constructing the key based on the outputs available.
• Users can expect a more consistent handling of outputs, which may, in turn, enhance the usability of YOLOv5 for various computer vision tasks including segmentation.
• There might be a performance improvement in scenarios where segmentation model outputs are used, due to a more direct access to the required output tensors. 🚀