feat: Add LLaDA 1.0/2.0 Diffusion Model Support

[akkikiki]

Mostly vibe-coded to see if we should resume the previous threads/PRs on the diffusion lm threads.
CC: @awni @Goekdeniz-Guelmez @cavit99

Happy to contribute on top of #14 and/or #270 if that's appropriate but any suggestions?

[kernelpool]

There's also #560 😀

[akkikiki]

@kernelpool Obviously, I missed it 😀
Any plans for merging it soon?
There's more to build on it in addition to supporting models (e.g., kv-caching for block diffusion), and since SGLang is fully investing on supporting it at sgl-project/sglang#14199, I assume that there will be enough tractions?

[Goekdeniz-Guelmez]

@akkikiki keep this pr open, the reason is that diffusion models need a careful, future-proof integration into mlx-lm. Eventually it will be merged, but will take its time. Keep this open and when the time is right, its getting merged :D