[enhancement]: Prompt enhancer as a standalone model type (ERNIE-Image) #9738

Is there an existing issue for this?

• I have searched the existing issues

What should this feature add?

Give ModelType.PromptEnhancer a config and a loader so ERNIE-Image's prompt enhancer installs as its own model.
Today the type exists without either, and the enhancer is only usable from inside the Diffusers pipeline, so a
single-file or GGUF ERNIE-Image install has no prompt enhancer.

Notes from an earlier investigation:

• The enhancer has the same geometry as the Ministral 3B encoder; the distinguishing signal is the missing Pixtral
  vision tower (only the encoder carries it). Today the encoder probe refuses the enhancer by name.
• Ministral3ForCausalLM's lm_head is tied and absent from the released file; it must be tied after loading,
  or a randomly initialised head generates nonsense without failing.
• apply_chat_template works on the Tekken tokenizer backend.
• Metadata: keep the typed prompt as positive_prompt and store the enhanced text in a separate field.

Alternatives

Use the Diffusers pipeline, which carries the enhancer.

Additional Content

Needs an optional prompt_enhancer_model on the ERNIE-Image loader node and a Generate slot.

The full matrix of what loads today is in #9726 (Model Format Support page).