🚨 fix(SigLip): remove spurious exclusion of first vision output token

[transmissions11]

Looks like the first token is spuriously excluded from the average pooling because this code was copy & pasted from modeling_clip, which does prepend a [CLS] token. However SigLip has no such special token, so it would seem we are needlessly excluding a token from being included?

p.s. why not use the pooled_output from the model's map head vs averaging the last hidden states?

[NielsRogge]

There's another PR which suggested to use the pooled output rather than averaging the patch tokens: #30373.

The reason I went with the latter was because of this paper (follow-up paper by the ViT authors which showed that it works better than the CLS token). However averaging patch tokens might work better for ViTs pre-trained using (self-)supervised image classification. For contrastive models like CLIP and SigLIP, perhaps it makes more sense to use the pooled output. I'm happy to change it, although we would need to make it backwards compatible

[amyeroberts]

Thanks for fixing this!

Only thing we need to do is run the model slow tests. Pushing the following should trigger the necessary CI run:
git commit --allow-empty -m "[run-slow] siglip"