Add Dream 7B Diffusion Large Language Model Pipeline

[dg845]

What does this PR do?

This PR implements a pipeline for the Dream 7B diffusion large language model (blog post, weights and code, repo). Dream is a masked (discrete) diffusion model for text which claims to perform comparably to similarly sized SOTA autoregressive LLMs such as Qwen 2.5 7B on NLP tasks and have superior performance on planning tasks.

Fixes #12017.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@yiyixuxu
@a-r-r-o-w
@ntoxeg

[dg845]

@yiyixuxu @a-r-r-o-w the Dream 7B model uses a transformers-style custom tokenizer, which I believe is based on GPT2Tokenizer but with different pre-tokenization rules. Since this tokenizer is not in transformers, should I open a PR there to add it? And if so, do you think the Dream transformer should also be added to transformers? The original transformer implementation is also transformers-compatible.

[a-r-r-o-w]

In case of certain custom implementations, we try to implement and keep the relevant files within diffusers. Some examples I could quickly find are:

• https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/ltx/modeling_latent_upsampler.py
• https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/flux/modeling_flux.py
• https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/consisid/consisid_utils.py

Maybe in this case you could create a tokenizer_gpt.py file within the pipeline directory to use it? WDYT @yiyixuxu?

I don't think the model implementation can live in transformers if we're using it for diffusion sampling. For example, Cosmos 1.0 released with an autoregressive and diffusion version, but we have two different implementations and PRs for support it in both libraries. So, let's maintain it here :)