Add support for Llama-Guard

[agunapal]

Can we please add support for llama guard.

Llama Guard offers a unique solution and it makes sense to support it in torchtune.

I'll be happy to send a PR if there is a template for the changes needed

[agunapal]

cc @ebsmothers

[ebsmothers]

@agunapal I think the main changes should be supporting the model builder and the tokenizer.

For the model changes, you can see the builders for our other Llama 3.2 models here. You should be able to define a version of the general 3.2 builder with tie_word_embeddings=False (I think the RoPE scaling may also be different?).

For the tokenizer, most things look the same (except eos_token and maybe the template?). You should be able to reuse Llama3Tokenizer in a builder like this, just change the EOS token as needed. I'm not sure about the difference with the template, but seems there was an open question around the usage of .strip() (see #2237), so we may need to sort that out.

[ebsmothers]

(Update based on chat: this is all assuming Llama Guard 3.2 1B. For the 8B Llama Guard model the process should be similar but with 3.1 instead of 3.2.)