Fix setting attention for multimodal models

[zucchini-nlp]

What does this PR do?

Fixes setting attention implementation in multimodals as a dict. Currently it fails because self._attn_implementation is not defined at the point when we try to get it. We need to set attention to None if the key is not found in dict, which is the default attention

[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.