Remove hook torch.nan_to_num(x)

[glenn-jocher]

Observed erratic training behavior (green line) with the nan_to_num hook (introduced in #8598) in classifier branch. I'm going to remove it from master.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improvement to the training stability by adjusting layer freeze behavior in the YOLOv5 model.

📊 Key Changes

• The hook that converts NaN (not a number) values to zero during training has been commented out.

🎯 Purpose & Impact

• Purpose: To address and possibly resolve erratic training results that may be caused by the previous method of handling NaN values.
• Impact: Users should experience more stable and reliable training, although they might need to monitor the training process for NaN values which are no longer being automatically converted to zeros.