Implementation of Git-Rebasin to Huggingface models

### Feature request Hi, I am not sure if this is even feasible/appropriate request in Huggingface. Recently, there was an ICLR paper titled "Git Rebasin" (https://arxiv.org/abs/2209.04836) which demonstrates a method to effectively merge models. But the authors have implemented it only for plain NN and Resnet kind of architectures. It would be really great if this feature is supported for Huggingface models as there'll be lots of instances where we might need to merge the models and a naive way of adding is not the best way. ### Motivation This feature is helpful especially in federated learning where 2 different people train 2 models separately and would like to combine their models. In academic circles as well, this is very much useful as there'll be lots of instances where effective model merging is critical! ### Your contribution I've raised PR's in subsequent Github repos 1. https://github.com/samuela/git-re-basin/issues/13#issue-1910375545 2. https://github.com/themrzmaster/git-re-basin-pytorch/issues/8#issue-1910392068 Not sure if this is feasible based on some of the previous discussions as the paper is based on "Permutation Invariance" of model weights and an architecture like Llama might have various possibilities thus difficulty in the method. I would be happy to contribute if there's some help!