[enhancement]: MiniMax H3 GGUF transformers #9727

Is there an existing issue for this?

• I have searched the existing issues

What should this feature add?

Load community GGUF builds of the MiniMax H3 transformer. Today installation is refused in
invokeai/backend/model_manager/configs/main.py with "GGUF-quantized MiniMax H3 checkpoints are not supported yet";
only Diffusers folders, bf16 single files and Comfy int8_tensorwise single files load.

Alternatives

Use the int8 single file (already supported) or the bf16 file.

Additional Content

• Demand on Hugging Face: unsloth/MiniMax-H3-GGUF (~1.7M downloads), Abiray/MiniMax-H3-GGUF,
  Abiray/MiniMax-H3-Pruned-GGUF, realrebelai/MiniMax-H3_GGUFs.
• Diffusers 0.41.0 adds single-file loading for MiniMax-H3 (Add single file support for Minimax H3 huggingface/diffusers#14839) and Krea 2
  (Add single file support for Krea 2 huggingface/diffusers#14914). InvokeAI pins diffusers==0.40.0 and vendors the MiniMax-H3 code, so the bump
  and its key mapping are worth evaluating here, both for the single file and for the GGUF key layout.
• Needs to handle the full and the pruned variant (the pruned flag lives on the config).
• The pattern from feat(ideogram4): load community GGUF transformers #9708 (Ideogram 4 GGUF) and feat(ernie-image): load GGUF transformers #9725 (ERNIE-Image GGUF) applies: a Main_GGUF_*_Config fingerprinted by keys (GGUF headers often name another architecture), a loader that keeps quantized Linears packed and unpacks the rest with unpack_ggml_at_load, and the denoise node's peak_dequant_transient_bytes, which already counts GGUF layers (fix(model-cache): reserve the GGUF dequantization transient in every node #9709).

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