Add support for visheratin/mexma-siglip2

[maxlund]

Claims to achieve new SOTA results on the multilingual retrieval benchmark Crossmodal-3600. See https://huggingface.co/visheratin/mexma-siglip2

[rwightman]

@maxlund looks like it might be possible to add with a model config and no code changes, though not quite sure about the text pooling https://huggingface.co/visheratin/mexma-siglip2/blob/cc7d73c89f452514ec3b996528543d026ad58f72/mexma_siglip.py#L70-L76

It's an XLMRoberta text enc + SigLIP2 image enc

Though I don't have time to do it so would need a contribution.

[maxlund]

@maxlund looks like it might be possible to add with a model config and no code changes, though not quite sure about the text pooling https://huggingface.co/visheratin/mexma-siglip2/blob/cc7d73c89f452514ec3b996528543d026ad58f72/mexma_siglip.py#L70-L76

It's an XLMRoberta text enc + SigLIP2 image enc

Though I don't have time to do it so would need a contribution.

For our purposes I can just use the transformers lib for now, too many things taking prio on the TODO list. But I'll hopefully have time the line to add a PR for adding support in open_clip as well, since it would make our codebase cleaner too.

Should also upate the benchmark results in e.g. https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_retrieval_results.csv, we have new SoTA models supported in the open_clip repo by now :)