Add CLIPS to open_clip

[Yanqing0327]

Add the CLIPS model, which is trained on public Recap-DataComp-1B.
Performance:

[rwightman]

@Yanqing0327 there's already a config based mechanism to override arguments to the normalization layer,

In the image and vision configs you can add

        "norm_kwargs": {
            "eps":  1e-6
        }

I'm also not quite clear on what the extra logic in the different tokenizer path involves, is there no way to use the existing HFTokenizer wrapper with some minimal pre/post processing that can be enabled via config?

[Yanqing0327]

Thank you so much for your help! I’ve already added the eps to the config as you suggested. However, for the tokenizer part, I haven’t found a cleaner implementation yet.