Feature: add option global_pool='max' to VisionTransformer

[fcossio]

Most of the CNNs models have a max global pooling option. I would like to extend ViT to have this option.

[rwightman]

@fcossio better to use amax for this... but before adding, I'd like some signal/indication that it's worth the LoC. Have you trained models from scratch or fine-tuned with the swap and shown noteworthy results vs avg pool?

[fcossio]

For high resolution images where the foreground has a small area, fine-tuning with max pooling worked slightly better both in target metrics and interpretability. I would like to contribute these LoC so that I don't need to keep a custom branch while at the same time it might benefit someone else.

Will be happy to edit the PR to amax if you think the PR is worth merging

[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.