Fix Ideogram4 full-model save: apply diffusers-to-original qkv fusion

[dxqb]

Summary

• IdeogramModelSaver.__save_safetensors saved the raw diffusers-split model.transformer.state_dict() directly, instead of routing it through model.checkpoint_diffusers_to_original() like every other fusing model's saver does (Flux, Flux2, Chroma, HunyuanVideo, PixArt-Alpha, SD3, Krea2, Anima).
• Ideogram4's native/ComfyUI checkpoint format uses a fused qkv nn.Linear per block — this isn't a Comfy-only quirk (unlike Z-Image, where fused qkv really is Comfy-side convenience); the real upstream ideogram-oss/ideogram4 architecture has a genuine fused self.qkv, and diffusers is the one that re-expresses it as split to_q/to_k/to_v for its own API. IdeogramModel already declares fusion_groups() and diffusers_to_original() correctly — the saver just never applied them.
• Symptom: a saved ORIGINAL_TRANSFORMER full-finetune checkpoint failed to load in ComfyUI:

  unet missing: ['layers.0.attention.qkv.weight', 'layers.0.attention.o.weight', ...]
  unet unexpected: ['layers.0.attention.to_q.weight', 'layers.0.attention.to_out.0.weight', ...]
  

Test plan

• Loaded the previously-saved (broken) full-finetune checkpoint and ran it through model.checkpoint_diffusers_to_original() + convert(): confirmed the output has fused attention.qkv/attention.o keys for all 34 blocks and zero remaining split to_q/to_k/to_v/to_out keys.

Drafted by Claude