gguf-py: preserve MoE size labels for mmproj metadata

[ooovenenoso]

Fixes #23448.

For mmproj exports, total_params refers to the projector, while the parsed Hugging Face model id still refers to the parent model. This preserves the parent model size label for mmproj metadata so Gemma 4 projector exports do not turn the parent size into finetune text.

Covered cases:

• google/gemma-4-26B-A4B-it with projector-sized total_params keeps general.size_label = 26B-A4B and general.finetune = it
• google/gemma-4-31B-it with projector-sized total_params keeps general.size_label = 31B and general.finetune = it

Tested with:

PYTHONPATH=gguf-py python -m pytest gguf-py/tests/test_metadata.py -q
ruff check gguf-py/gguf/metadata.py gguf-py/tests/test_metadata.py conversion/base.py

AI assistance was used for code exploration and validation. I reviewed the change and understand the metadata parsing logic.

[dfriehs]

I don't think this is the right choice here, the mmproj for gemma-4-26B-A4B-it is not even close to 4B parameters. The ~500M parameters were simply close enough in this case to allow A4B to be used as the size label.

I think the mmproj should either always or never take the size label of the model it belongs to, no matter if it's number of parameters would allow part of the name as a label.

[ooovenenoso]

Updated the PR based on the mmproj feedback:

• The code now treats mmproj parsing separately: total_params is the projector size, so the parent model size label is preserved from the model id.
• The test now uses a projector-sized value (576M) instead of 4B.
• Added coverage for both reported Gemma 4 cases:
  • google/gemma-4-26B-A4B-it -> size_label = 26B-A4B, finetune = it
  • google/gemma-4-31B-it -> size_label = 31B, finetune = it

Verified locally:

PYTHONPATH=gguf-py python -m pytest gguf-py/tests/test_metadata.py -q
ruff check gguf-py/gguf/metadata.py gguf-py/tests/test_metadata.py conversion/base.py