[bug]: MM2 downloads extra files from diffusers repos

[psychedelicious]

[bug]: MM2 downloads extra files from diffusers repos

[psychedelicious]

On closer inspection I see we are filtering the files in invokeai/backend/model_manager/util/select_hf_files.py, and there is a note about how we prefer .safetensors files, so never mind mine comment about downloading .bin files.

The remaining issue is downloading extra files. This occurs because the safetensors and bin files have different names, despite representing the same weights. The filter function sees them as different files instead of different versions of the same data.

For example, text_encoder/model.fp16.safetensors" and text_encoder/pytorch_model.fp16.bin` are the same, but with different base names - "model" vs "pytorch_model".

I don't see any simple way to fix this given the very loosey-goosey structure of the diffusers format.

[psychedelicious]

Figured out a fix in #5846.

We only want the single "best" weights file for each submodel type/directory. We prefer safetensors and exact variant matches. To choose the right one, we can score each file and take the best one.