[add] add fp16 training

[Muyun99]

Hello, chuanguang

It`s a PR about fp16 training. I use this code to run your CIRKD on four RTX 3090 without hyperparameter changes. The performance is still aligned with faster training speeds and reduced GPU memory.

I implemented fp16 in a straightforward way with if-else style. If you have any suggestions on code style please contact me. BTW, I just started my internship at Horizon, you can find me on feishu.

Looking forward to your reply. Thanks!

Best Regards! Yun

[winycg]

Hello Yu Thanks for your great contributions！ I am very glad to merge your fp16 training code to this codebase. It seems that the added code lengthens the original training files and may impact the simplicity of the core code. I hope you can create a single training file to contain your fp16 training code, such as train_kd_fp16.py. Thanks for your contributions again!

Best Regards! Chuanguang