Update cls bias init

[glenn-jocher]

Increased numerical precision. Returns 1.0 probability for single-class datasets now. Addresses #5357

torch.sigmoid(torch.tensor([math.log(0.6 / (1 - 0.999999))]))
Out[19]: tensor([1.0000])

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Adjusted initial bias in YOLO object detection model.

📊 Key Changes

• Modified the calculation for class probability bias initialization in the object detection layer.

🎯 Purpose & Impact

• 🎯 Purpose: To fine-tune the initial bias calculation for better stability in class prediction, preventing overconfidence in the less frequent classes.
• 💥 Impact: Should slightly improve the robustness and performance of the model, particularly in datasets with class imbalance. Users may notice better class probability estimates during training and inference.