Adding CrossMAE

### Model description Hey, the recently released [CrossMAE](https://crossmae.github.io/) seems like it would be a nice addition to transformers. Basically the model improves MAE by using Cross-Attention instead of Self-Attention on the tokens and thereby decreasing the needed FLOPS quite significantly. At the same time it seems to be able to keep the performance of MAE or even improve it a bit. Maybe there are already plans of integrating it @NielsRogge ? ### Open source status - [X] The model implementation is available - [X] The model weights are available ### Provide useful links for the implementation Project Page: https://crossmae.github.io/ GitHub Repo: https://github.com/TonyLianLong/CrossMAE Paper: https://arxiv.org/pdf/2401.14391.pdf