Add --half support for FP16 CoreML exports with

[Cedric-Perauer]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved CoreML model export with optional quantization.

📊 Key Changes

• Added int8 and half arguments to export_coreml() function enabling weight quantization.
• Wrapped the CoreML weight quantization code with a check for macOS, as quantization is only supported on that OS.
• Updated the run() function to conditionally enable FP16 (half) mode based on whether CoreML export is enabled.
• Conditionally run weight quantization depending on the presence of int8 or half flags and the operating system.

🎯 Purpose & Impact

• 🔍 Purpose: This update aims to provide users with the ability to export CoreML models with reduced size and potentially improved performance on devices by using weight quantization (either 8-bit or 16-bit precision).
• ⚡ Impact: Users exporting YOLOv5 models to CoreML format can now choose to quantize their models, leading to more efficient models especially suitable for mobile and edge devices. However, the quantization features are limited to macOS users due to platform restrictions.

[glenn-jocher]

@Cedric-Perauer PR is merged. Thank you for your contributions to YOLOv5 🚀 and Vision AI ⭐