Fix position encodings for Pixtral

[ameroyer]

For Pixtral models, the position encodings are resampled on-the-fly when dealing with images which are smaller than the default model's config (1024x1024px images, or 64x64 patches of size 16px x 16px). Doing so seems to slightly improve generations (in particular when it comes to the number of objects).

• link to ref in Huggingface code
• testing the fix : mainly looking at the model outputs for different image size

[LaurentMazare]

Thanks, you might want to look at the clippy failure in the CI :)

[ameroyer]

Indeed, clippy again, thanks it's fixed now !

• Added a minor fix: The default hidden activation for the vision config has been corrected from GeLU to SiLU in the HF repo

[LaurentMazare]

Thanks!