[tests] make test_model_parallelism device-agnostic

[faaany]

What does this PR do?

As the title suggests, test_model_parallelism is not GPU-specific and could also work on other accelerators, e.g. XPU device. Below are the before and after the fix for pytest -rA tests/models/bart/test_modeling_bart.py::BartModelTest::test_model_parallelism:
BEFORE

================================================ short test summary info ================================================
FAILED tests/models/bart/test_modeling_bart.py::BartModelTest::test_model_parallelism - AssertionError: device(type='xpu', index=0) != device(type='cuda', index=0)
============================================= 1 failed, 4 warnings in 3.75s =============================================

AFTER

================================================ short test summary info ================================================
PASSED tests/models/bart/test_modeling_bart.py::BartModelTest::test_model_parallelism
============================================= 1 passed, 4 warnings in 5.96s =============================================

Pls help revie this PR, thx! @amyeroberts and @ydshieh

[ydshieh]

Good from my side, thanks. Let's maybe wait a remark the our accelerate experts before merge.

[SunMarc]

Thanks for the fix ! Left a comment

[amyeroberts]

Thanks!