[enhancement]: Ministral 3B GGUF text encoder (ERNIE-Image) #9731

Is there an existing issue for this?

• I have searched the existing issues

What should this feature add?

Load GGUF builds of ERNIE-Image's Ministral 3B text encoder. Today configs/mistral_encoder.py refuses them
("Ministral 3B GGUF encoders are not supported"): the Mistral GGUF loader builds a MistralModel with RoPE read
from GGUF metadata, while Ministral 3B needs Ministral3Model (YaRN scaling and a position-dependent attention
scale). Packed nvfp4 Ministral builds are refused as well.

Alternatives

Use the 7.7 GB safetensors encoder (the ERNIE-Image starters install it).

Additional Content

• Demand: moritalous/Ministral-3-3B-Base-ERNIE-Image-TE-sdcpp-GGUF, ponpoke/ERNIE-Image-Abliterated-GGUF
  (text encoder files).
• With ERNIE-Image GGUF transformers (feat(ernie-image): load GGUF transformers #9725) at ~5 GB, the encoder is now the largest part of a GGUF install.

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