Refactor out-of-place Detect() for reduced ops

[glenn-jocher]

Reduced layers in ONNX, CoreML etc. exports

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced coordinate calculation in YOLOv5 model.

📊 Key Changes

• Updated the formula used to calculate the x, y coordinates in bbox prediction.
• Modified the torch.cat axis parameter in the AWS Inferentia compatible code path.

🎯 Purpose & Impact

• Ensures more accurate spatial transformations for bbox predictions, potentially leading to better model performance.
• Correction of the tensor concatenation process ensures proper alignment of the bounding box (bbox) attributes, which could improve inference results on AWS Inferentia-powered devices.