Fix text.transformer.embeddings.position_ids key error

[EIFY]

This PR is to fix #584.

After transformers==4.31, certain text transformers no longer expect text.transformer.embeddings.position_ids (huggingface/transformers#24505). The recommendation is to use from_pretrained (huggingface/transformers#25330 (comment)) but that's not an option for loading a module. So, I just explicitly identify this inconsistency and edit state_dict if necessary. Tested with both transformers==4.30.2 and transformers==4.31.

[usuyama]

Thank you!

Hope this PR fixes this issue in BiomedCLIP https://huggingface.co/microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224/discussions/9

[Adenialzz]

transformer==4.28.0 failed here with the newest open_clip version

[EIFY]

@Adenialzz I can't replicate, loading eva_giant_patch14_224 with the latest main works for me. In fact, I believe transformers==4.28.0 isn't directly used at all for loading the model for inference. This PR just makes sure that the code can load newer checkpoints from transformers>=4.31.