More RLHF algorithms in the implementation

[WayXG]

I saw the choice of the loss type indicating that several other loss functions can be used like hinge, ipo, raft ...

I am wondering whether we only need to modify the loss choice and do not need to modify other parts of the codes.

[hendrydong]

Yes. That is also our purpose.

We leave the loss type api for users and researchers to develop other loss functions beyond KL-induced ones.

We hope the community can find better recipes and solutions for the whole pipeline.