Add a sampling animation example for discrete diffusion

[kashif]

Adds examples/discrete_diffusion/animate_sampling.py, a small tool to visualize how discrete diffusion language models write text.

It captures the intermediate canvas of every denoising step through each pipeline's callback_on_step_end (no changes to the pipelines) and renders it as a self-contained HTML animation or an animated GIF. The goal is to make the two families easy to compare: LLaDA2 reveals text from [MASK] slots, while DiffusionGemma sharpens a mask-free canvas of random tokens block by block.

For DiffusionGemma you can pick any of the three supported schedulers (entropy_bound, discrete_ddim, block_refinement) to compare them on the same prompt and seed. Trajectories with different step counts are resampled onto a shared timeline so the panels stay in sync, and the highlight colors are color-blind and grayscale safe.

The README has a new section with capture and render commands. The script is also runnable with uv.

[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.