feat: add LLaDA-Image support

[fszontagh]

Summary

Adds text-to-image and instruction-guided editing for
LLaDA-Image (Apache-2.0). The denoiser is a
Lumina2/Z-Image-style NextDiT conditioned by a LLaDA2-MoE diffusion-LLM text encoder, and it
reuses the Flux.2 VAE.

Most of it reuses what is already here:

• The DiT instantiates ZImage::JointTransformerBlock / FinalLayer; the hyperparameters are
  identical. Z-Image's blocks gained two defaulted parameters, norm_elementwise_affine
  (LLaDA's norms are parameter-free) and split_qkv (its checkpoint keeps q/k/v separate).
• The text encoder is a new LLMArch::LLADA2_MOE on the existing MoE path, adding
  grouped-sigmoid routing with an expert bias, a shared expert, partial rotary and
  bidirectional attention.
• New: QueryFormer, text projection, and SigVQ (editing only, constructed only when those
  weights are present), plus the llada_image reference-image preset and sigma schedule.

Editing runs the clean reference and the noisy target in one sequence. Because adaLN is a
linear map of the timestep embedding, feeding a per-token embedding selects the right
modulation exactly, so no new modulation machinery was needed.

One change outside the model: tensor_should_be_converted now skips *_pad_token.
LLaDA-Image stores those around 7.7e24, so f16 storage and q8_0's f16 block scale both
overflow to inf, and every GGUF I converted rendered blank white until this was added.

Docs in docs/llada_image.md; rows added to README.md and docs/edit.md. The connector
components ship in separate directories, so the docs show how to merge them into the one file
--embeddings-connectors expects, and link pre-merged copies.

Related Issue / Discussion

Closes #1937

Additional Information

Verified on an RTX 3060 12GB, CUDA, with LLaDA-Image-Turbo int8_tensorwise and with q8_0 and
f16 GGUFs converted by -M convert, against a Q4_K text encoder. No offload flags are needed.

512x512 / 768x768 / 1024x1024	10.5 s / 22.3 s / 35.5 s
1024x576, 576x1024	22.4 s
editing, 512x512	107 s

assets/llada_image/example.png is "a lovely cat holding a sign says 'llada.cpp'";
edit_example.png is that image with "change the sign text to 'sd.cpp'".

Segmented execution was checked against unconstrained execution at 512x512: --max-vram 6,
--max-vram 4, --max-vram 3 and --offload-to-cpu all produce byte-identical output, so the
roughly 16 GB of weights run in a 3 GB budget.

Editing packs reference and target in one sequence, so it needs about twice the tokens of
text-to-image at the same size; at 1024x1024 on 12 GB it needs --diffusion-fa.

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. Tested with
int8_tensorwise and GGUF quantizations only.

Checklist

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

[leejet]

Please put merge and vocab on a single line, and remove the helper scripts.