[gpt2-int8] Add gpt2-xl int8 test

[younesbelkada]

What does this PR do?

Addresses: #24504 (comment)

Currently the test is failing because of the following reason (that I need to explore and as discussed offline @sgugger):
1- If a buffer is defined as persistent=False in the modeling file
2- and not present in the state_dict
The dispatch_model seems to fail as I face device mismatch issue. I think a potential fix needs to be addressed on accelerate
Facing the same issue for blip2 int8 tests, as the cos_cached and sin_cached are defined as buffers with persistent=False and not present in the state dict.

Seems that the issue does not happen before 5791d949ff93733c102461ba89c8310745a3fa79 on accelerate

cc @sgugger

EDIT: huggingface/accelerate#1660 should fix the issue I am facing

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[younesbelkada]

Can confirm all the tests pass now (with the accelerate PR mentioned above)

[sgugger]

Thanks for adding this test!