Fix flaky test_beam_search_low_memory

[ydshieh]

What does this PR do?

test_beam_search_low_memory is very flaky for Emu3Vision2TextModelTest.

The 3 fixes in this PR are all necessary to have 0 failure in a suite of 100 runs.

• call set_xxx_less_flaky methods
• avoid generate image tokens
• compare with _check_similar_generate_outputs

The final version run 500 times without failure.

[ydshieh]

This is still required, otherwise 1% of chance to fail for Emu3Vision2TextModelTest

[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.

[zucchini-nlp]

LGTM, thanks!

[ydshieh]

tests/models/mobilenet_v1/test_modeling_mobilenet_v1.py::MobileNetV1ModelTest::test_batching_equivalence - AssertionError: tensor(False) is not true : Batched and Single row outputs are not equal in MobileNetV1ForImageClassification for key=logits. Difference=0.48479270935058594.

This is a bit nasty but irrelevant to this PR. Will merge and try to fix the above one next week.

[zucchini-nlp]

@ydshieh Cool, feel free to close #29516 after fixing that. We had a small tracker for flaky vision model with that test

[ydshieh]

Thanks. The tracker is more for batching tests, but I will check with #35564 next week