Implementing exaone3.5

[KareemMusleh]

This is my first attempt at an implementation of exaone into unsloth, it was requested in this issue.

I didn't want to implement exaone into a separate model class because exaone follows the llama architecture as was discussed in this issue. As of now I am having problems with the from_pretrained function when using config and state_dict. I've already opened an issue in transformers about it. I'll try to solve that problem soon.

[qingy1337]

@KareemMusleh Thanks for the effort! I was just curious of the general process you go through to implement something like this. Is it different for every model, and how so?

[KareemMusleh]

@qingy1337 I was thinking about creating a writeup for this (assuming this gets accepted), as this is my first pull request to a large public repo.

But the gist of it is this:
I started out by googling how other people have implemented exaone into their codebase, and I came up with 3 possible solutions:

1. Write the exaone class from scratch like they did in vLLM
2. Upload the exaone models to huggingface as llama models, as was done by maywell
3. Map exaone to llama, which is what I did, as was recommended in this issue

This is different than previous model integrations because exaone has the exact same architecture as llama

[KareemMusleh]

For this PR to be ready the all attention refactor fix PR should be merged first. Because it uses the latest transformers