Add PYTEST_TIMEOUT for CircleCI test jobs

[ydshieh]

⚠️ Before merging, we need to run the full tests on CircleCI to see if there are slower tests that will fail and decide what to do with them.

What does this PR do?

Add PYTEST_TIMEOUT: 30 for CircleCI jobs:

environment:
    ...
    PYTEST_TIMEOUT: 30

The main goal is to avoid CircleCI's default 10 minute timeout that cancels the jobs. Also with this PR, we can see clearly which test (s) timeout.

[HuggingFaceDocBuilderDev]

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

[sgugger]

I like the idea of getting a hard error instead of silently getting new tests that slow down the CI by quite a lot! Now we just ahve to get to all tests passing below that threshold 😅
For the examples, you can authorize 60s before timing out as those are end-to-end small training so take more time.

[ydshieh]

Currently set to PYTEST_TIMEOUT: 120. As mentioned on Slack, tests sometimes get much longer to run. For example,

test_modeling_data2vec_audio.py::Data2VecAudioModelTest::test_mask_time_prob_ctc

44.16s call 
37.64s call     
12.46s call     
12.60s call 

and

test_modeling_plbart.py::PLBartBaseIntegrationTest::test_base_generate

48.18  call
11.20s call 
11.53s call

This makes it difficult to determine a good threshold that won't be flaky.

Current longest 2 tests

(observed on a CircleCI workflow run):

73.08s call     
test_pipelines_image_segmentation.py::ImageSegmentationPipelineTests::test_pt_DetrConfig_DetrForSegmentation_notokenizer_DetrFeatureExtractor

65.33s call
longt5/test_modeling_flax_longt5.py::FlaxLongT5ModelTest::test_jit_compilation

[ydshieh]

Will revert this line once this PR being approved

[LysandreJik]

LGTM!

[ydshieh]

Reverted the change in setup.py (was for running the full tests). The final timeout limit is 2 minutes (to avoid flaky failures).
I will merge this PR today unless @sgugger has a different opinion.