FA2 support for Aria

[molbap]

Currently, Aria does not support FA2 in its modeling code:

transformers/src/transformers/models/aria/modeling_aria.py

Line 656 in 2fa876d

_supports_flash_attn_2 = False

However the reference checkpoint for the Aria model does set attn_implementation to FA2: https://huggingface.co/rhymes-ai/Aria/blob/104c6548d7da08fc3a30c4232d35cc9fb3239942/config.json#L36

So the modeling code on transformers side should be updated to support it. It currently causes errors on the CI and I think it's due to this mismatch.

cc @aymeric-roucher @zucchini-nlp

[zucchini-nlp]

I don't have write access to the repo, but I guess there should be a colab in slack to ping the authors