Fix AMP check tolerance

[glenn-jocher]

Adjust to 5%, fixes failing Colab AMP check with V100 (1.5% different) with 200% safety margin.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced mixed-precision inference validation in YOLOv5's utility functions.

📊 Key Changes

• Changed output format from bounding box corners (xyxy) to normalized bounding box format (xywhn) during inference checks.
• Tightened the absolute tolerance from 1.0 pixel to 5% (0.05) when comparing FP32 and AMP inference results.

🎯 Purpose & Impact

• Purpose: To improve the reliability of mixed-precision (AMP) inference checks by using a format (xywhn) that's typically used for evaluating object detection models and making the validation process more stringent.
• Impact: This change leads to more robust verification of whether mixed-precision inference is functioning correctly without significant deviation from full-precision (FP32) results, potentially improving overall model accuracy and stability. It can also benefit developers and researchers by providing a more reliable assessment method when implementing mixed-precision support in their projects. ⚙️📈