Add plain PyTorch loader for the CSD checkpoint

[fahmeed-nabi]

Summary

Adds a plain PyTorch loader for the released CSD checkpoint, answering #8 ("How to load the model as a PyTorch module?").

The checkpoint has no transformers AutoModel support and config.json carries no architecture info ({"model_type": "custom"}), so it can't be loaded with from_pretrained. models/csd_clip.py reconstructs the architecture directly from the checkpoint's state_dict key names and shapes:

• The backbone is a stock CLIP ViT-L/14 (quickgelu) visual transformer, matching OpenAI's original CLIP activation since training started from OpenAI's CLIP weights rather than open_clip's plain-GELU variant.
• The backbone's own contrastive proj is unused (no proj key in the checkpoint), and the raw pooled feature feeds forward instead.
• last_layer_style / last_layer_content are separate (1024, 768) matrices applied via matrix multiply, not nn.Linear (the checkpoint stores them as raw (in, out) matrices, not nn.Linear.weight's (out, in) shape).

Loading with strict=True against this class produces zero missing and zero unexpected keys against the released weights.

Also adds a short README section with a runnable example (load the checkpoint, embed an image, get back style/content embeddings).

Why

This has come up a few times (#8, and implicitly in how people are using the HF checkpoint) with no canonical answer in the repo. This closes that gap with a minimal, dependency-light loader that doesn't require guessing the architecture from the state_dict, which is what we had to do to use CSD in our own project.

Testing

• Verified strict=True state_dict load produces zero missing / zero unexpected keys against tomg-group-umd/CSD-ViT-L.
• Ran the README example end-to-end: loads the checkpoint, embeds an image, produces style and content embeddings of the expected shape.

Closes #8.