example: add patch dropout to open_clip

[iejMac]

original PR: mlfoundations/open_clip#272

[iejMac]

openai can you summarize this PR?

[github-actions]

This pull request adds a patch dropout function to the CLIP Vision model, which is a transformer model used for vision-language tasks. The patch dropout function helps improve model performance by randomly dropping visual tokens prior to the transformer layer. Additionally, the init_values parameter is set to 1e-5, and a LS init value of None is added to the CLIPVisionCfg class. Possible improvements could include adding more efficient pooling layers such as AdaptiveAvgPool2d, and adding a new parameter to the patch dropout function to increase the amount of tokens dropped.