MAINT enable verbose pytest on Azure to debug #MercuryRetrograde stalled builds

[ogrisel]

Trying to understand why some builds stall on Azure Pipelines despite the merge of #22302 that disabled pytest-xdist.

Edit: example stalled test after the merge of #22302:

• Linux_Nightly pylatest_pip_scipy_dev at 43% apparently in neighbors/tests/test_neighbors.py: https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=37402&view=logs&j=dfe99b15-50db-5d7b-b1e9-4105c42527cf&t=ef785ae2-496b-5b02-9f0e-07a6c3ab3081
• Linux pylatest_pip_openblas_pandas at 97% in utils/tests/test_optimize.py: https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=37396&view=logs&j=78a0bf4f-79e5-5387-94ec-13e67d216d6e&t=f1857171-4a53-55c7-3ab5-90acfe091baa
• macOS pylatest_conda_forge_mkl at 97% in utils/tests/test_optimize.py: https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=37396&view=logs&j=97641769-79fb-5590-9088-a30ce9b850b9&t=4745baa1-36b5-56c8-9a8e-6480742db1a6
• Linux pylatest_pip_openblas_pandas at 97% inutils/tests/test_optimize.py: https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=37404&view=logs&j=78a0bf4f-79e5-5387-94ec-13e67d216d6e&t=f1857171-4a53-55c7-3ab5-90acfe091baa
• macOS pylatest_conda_forge_mkl at 47% in neighbors/tests/test_neighbors.py https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=37404&view=logs&j=97641769-79fb-5590-9088-a30ce9b850b9&t=4745baa1-36b5-56c8-9a8e-6480742db1a6

Stalled builds from this PR with verbosity enabled:

• Linux_Runs pylatest_conda_forge_mkl after utils/tests/test_parallel.py::test_configuration_passes_through_to_joblib[multiprocessing-1] PASSED [ 97%] https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=37403&view=logs&j=dde5042c-7464-5d47-9507-31bdd2ee0a3a&t=4bd2dad8-62b3-5bf9-08a5-a9880c530c94
• Linux pylatest_pip_openblas_pandas after after neighbors/tests/test_neighbors.py::test_same_radius_neighbors_parallel[auto] PASSED [ 43%] https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=37403&view=logs&j=78a0bf4f-79e5-5387-94ec-13e67d216d6e
• macOS pylatest_conda_mkl_no_openmp after neighbors/tests/test_neighbors.py::test_knn_forcing_backend[kd_tree-testing] PASSED [ 47%] https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=37403&view=logs&j=e6d5b7c0-0dfd-5ddf-13d5-c71bebf56ce2&t=c500742d-7cbe-569e-8da5-94db9b4cd21e

[glemaitre]

I just saw that coverage has been updated from 6.2 to 6.3 on January 25.
It could be worth launching a build downgrading the version and with xdist just to check that this is not linked to this.

[ogrisel]

I just saw that coverage has been updated from 6.2 to 6.3 on January 25.
It could be worth to launch a build downgrading

Please feel free to do that in a parallel PR to avoid cancelling the current running jobs.

[glemaitre]

It seems that this is not a particular test.

[glemaitre]

I see the light...

nedbat/coveragepy#1310

[ogrisel]

Ok so based on the linked discussion, coverage 6.3 is no longer multiprocessing fork-safe. And indeed we have a few tests that try to use multiprocessing fork instead of loky and the verbose output of this PR seems to point to those tests (see the summary in the description).

I think this is a regression in coverage. +1 for pinning it for now. Let's see if #22311 passes.

[ogrisel]

Closing in favor of #22311.