[DiffusionGemma] Add DDIM and block refinement samplers

[kashif]

Adds two more samplers for DiffusionGemma so people can try different denoising strategies without writing their own.

• DiscreteDDIMSampler: samples each canvas position from the exact posterior of the uniform corruption process. Works well with a constant temperature and around 20 steps.
• BlockRefinementSampler: commits an even share of the canvas per step, most confident tokens first, with optional editing of committed tokens.

The sampler is picked through sampler_config, same as the existing EntropyBoundSampler. Added unit tests for both, parametrized the generate test over them, and documented them in the model page.

Tests: pytest tests/models/diffusion_gemma/test_generation_diffusion_gemma.py and test_modeling_diffusion_gemma.py -k test_generate pass on CPU.

AI assistance was used.

[kashif]

cc @gante and @sayakpaul

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[gante]

LGTM!

Out of curiosity, have you benchmarked any of the new samplers with diffusiongemma? :)

[kashif]

running benchmarks as i type

[gante]

It would be cool to measure both quality and tokens per forward! (I added tokens per forward as part of the output class in generate)

[gante]

P.S.: In the first few commits I considered writing a base class for all samplers, since they will almost certainly share the random canvas initialization 🤗

[github-actions]

CI Dashboard: View test results in Grafana

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: diffusion_gemma