feat: add LLaDA-Image support

[fszontagh]

Summary

Adds text-to-image and instruction-guided editing for LLaDA-Image (Apache-2.0): a 6B Lumina2/Z-Image-style NextDiT conditioned by a LLaDA2-MoE diffusion-LLM text encoder, reusing the Flux.2 VAE.

Most of it reuses what is already here. The DiT instantiates ZImage::JointTransformerBlock and FinalLayer with identical hyperparameters, and the text encoder is a new LLMArch::LLADA2_MOE on the existing MoE path. New are the QueryFormer, the text projection, and SigVQ for editing, plus the llada_image reference-image preset and sigma schedule. Editing runs the clean reference and the noisy target in one sequence; adaLN is linear in the timestep embedding, so feeding a per-token embedding selects the right modulation without new modulation machinery.

The LLaDA2 vocabulary is not embedded, so --tokenizer with the model's tokenizer.json is required, per the embedded-data allowlist in CONTRIBUTING.md.

One change outside the model: tensor_should_be_converted now skips *_pad_token. LLaDA-Image stores those around 7.7e24, which overflows both f16 storage and q8_0's f16 block scale, so every converted GGUF rendered blank white without it.

Docs in docs/llada_image.md; rows added to README.md and docs/edit.md.

Related Issue / Discussion

Closes #1937

Additional Information

Verified on an RTX 3060 12GB, CUDA. Both checkpoints work: LLaDA-Image-Turbo does text-to-image and editing at 512x512 and 1024x1024, and the 50-step base model at --steps 50 --cfg-scale 5 does both at 1024x1024. At 512x512 the base model's editing returns the reference almost unchanged, which the docs note.

--max-vram 6, --max-vram 4 and --max-vram 3 all produce byte-identical output to unconstrained execution.

Not implemented: generation_mode="vq", where the text encoder block-diffusion-decodes VQ tokens before diffusion. Text-to-image and editing do not use that path.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

[fszontagh]

Both done: vocab and merges are on a single line now, and the helper scripts are removed.

Since the merge script is gone, I converted the weights for both checkpoints and uploaded them, so the docs link ready-made files instead of asking people to build their own. Links are in docs/llada_image.md.

[fszontagh]

Rebased on master and dropped the embedded vocab, following #1974. LLaDA2 now needs --tokenizer with the model's tokenizer.json, and fails with a clear message if it's missing. Also fixed the encode() call for the new signature.

Output is pixel-identical to the old embedded-vocab build.

Updated the PR description too.