Add TinyCLIP

[gau-nernst]

Close #2025

Ref: https://github.com/microsoft/Cream/tree/main/TinyCLIP

Auto-pruned variants (63M, 45M, 22M) have irregular structure (different head dim at each layer) so it's not possible to integrate them easily.

@rwightman Also, TinyCLIP weights are in float16, so resample patch embed and pos embed will fail. Should I add a cast to float32 to these functions?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[rwightman]

@gau-nernst FYI, thinking about the naming here... the 39/40M model defs are actually the same as my 'medium' that I've been using a while (and have more weights pending release). Also in OpenCLIP https://github.com/mlfoundations/open_clip/blob/main/src/open_clip/model_configs/ViT-M-16.json

I also have a few embed=640, heads=10 weights lying around that I hadn't named yet but was calling medium+ in my refs.

So I'm going to have a moment to think about the naming...

[rwightman]

Ideas in slotting names within current vit sizing...

name - width, heads, depth
vit_tiny - 192, 3, 12
vit_xsmall, ? - 256, 4, 10
vit_small - 384, 6, 12
vit_medium - 512, 8, 12
vit_betwixt, vit_mediumw, vit_mediump, ? - 640, 10, 12 (w=wide, p=plus)
vit_base - 768, 12, 12
vit_large - 1024, 16, 24