fix fuyu device_map compatibility

[SunMarc]

What does this PR do ?

This PR fixes the multi-gpu issue that users were experiencing here #29848. This happened because this PR kinda reverted the changes I made to make it compatible with device_map here.
This is not very elegant since we need to modify the forward a bit, but this is an exception since we are dealing with the output of the vision model and the input of the language model.

[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.

[ArthurZucker]

Thanks 🤗

[grahamannett]

fyi this seems to actually still be broken as image_patches_indices can be wrong device for gather_continuous_embeddings

[grahamannett]

@ArthurZucker and @SunMarc seems like it might be helpful to specify a test on this and what sort of GPU setup the test has? Took me a few days to pinpoint but I believe I get an error when device_map='auto' for 8x GPUs but not on 2x (and smaller GPU's)?

Seems with more GPU's for multimodal models theres likely to be less chance the embeddings are on the same device for many HF models and not sure if that would be considered a bug or is specific to how infer_auto_device_map from PreTrainedModel. Have not looked too deeply into if other models are casting embeddings to same device or what the overhead on that is compared to having embeddings all on same device to start with (also seems this could be issue with FSDP as well from my understanding as transformer_auto_wrap_policy is used with decoder blocks?).