Fix the loss can be inf when training in FP16 without Pre-Trained model.

In this version, fix some errors.

1. Fix the lost prarms in pre-trained weight, such as yolov8n.bin.
2. Fix Attention blocks for yolov11 and yolov12.
3. Fix the bug in Detect module.
4. Use Imagick for image read or write to disk.

Add segment pretrained models.

Add pre-trained models