Add support for loading single file CLIPEmbedding models

[lstein]

Summary

We're starting to see fine-tuned CLIPEmbedding models for improved FLUX performance appearing in the wild. For example: https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14 . However, these fine-tunes are formatted as single "checkpoint"-style files rather than Transformers-compatible folders. This PR adds support for installing and loading these models.

Related Issues / Discussions

There is a problem with this implementation. The CLIP text embedder needs two models: the encoder and the tokenizer. When FLUX support was added to InvokeAI, these two models were grouped together under a single folder, and are treated as two submodels:

└── clip-vit-large-patch14
    ├── text_encoder
    │   ├── config.json
    │   └── model.safetensors
    └── tokenizer
        ├── merges.txt
        ├── special_tokens_map.json
        ├── tokenizer_config.json
        └── vocab.json

However, the single-file format contains just the text encoder and not the auxiliary files needed for the tokenizer. So as a workaround, when a CLIPEmbed single-file's tokenizer is requested I call CLIPTokenizer.from_pretrained() to download the tokenizer from the InvokeAI/clip-vit-large-patch14 HF repository. After it is downloaded, it is cached in the HuggingFace cache, so the next access does not require network. This is preferable to loading the tokenizer from the locally-installed clip model because (1) there is no guarantee that this has been installed previously; and (2) doing this would be incredibly ugly, and require the low-level loader to communicate with the high level model manager.

The main downside is that the first time the tokenizer is needed, the backend will hit the network, which is something we are trying to avoid. (see PR #6740 )

QA Instructions

Use the model manager tab to install one of the "HF" format CLIPTextModel models located at https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14. Try to render with it. The Tokenizer and TextEncoder should load and run successfully.

Merge Plan

Merge when approved.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)

[RahulVadisetty91]

You need to add the following lines to include the new model configurations.

    Annotated[SpandrelImageToImageConfig, SpandrelImageToImageConfig.get_tag()],
    Annotated[T2IAdapterConfig, T2IAdapterConfig.get_tag()],