Tests: upgrade test_eager_matches_sdpa_generate

[gante]

What does this PR do?

test_eager_matches_sdpa_generate has failed in our new failure reporting system (here, cc @ydshieh )

Having a look at the test, the cause for flakiness was clear: we are using random models with generate, and tiny perturbations can result in a different sampled token, causing generation to go in a different direction and ultimately failing the check (eager generate == sdpa generate).

This PR:

1. Moves the test to GenerationTesterMixin, as it calls generate
2. Adds logic to handle the flakiness and do the correct check. If the generation is different, check the logits in the first different token. Despite resulting in a different token, the logits should be nearly identical -- if they are, the test pass.
3. Remove most overwrites, which only existed to handle flakiness through is_flaky().

---

The following test commands were run:

1. RUN_SLOW=1 py.test tests/models/ -k test_eager_matches_sdpa_generate ✅
2. RUN_SLOW=1 py.test tests/models/gpt2/test_modeling_gpt2.py::GPT2ModelTest::test_eager_matches_sdpa_generate --flake-finder --flake-runs 500 ✅

[gante]

(this is mostly copy-paste, going to comment the sections that are changed)

[gante]

Uses self.prepare_config_and_inputs_for_generate() instead, which enables us to pass a dictionary of inputs to generate (better input control than simply using inputs_dict[model_class.main_input_name])

[gante]

Uses dictionaries -> more compact

[gante]

flakiness handling as explained in the PR header

[ArthurZucker]

Nice!

[ArthurZucker]

with #34282 we won't have to init 2 models.
Also this is memory hungry

[gante]

TBH we can do it one model at a time, going to change the test.

After Flex attention becomes the norm, we probably won't need this test

[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.

[ydshieh]

This is sooooo great! ❤️