`run_summarization` likely race condition creates flaky `test_tensorflow_examples`

[molbap]

As in title. See for instance https://app.circleci.com/pipelines/github/huggingface/transformers/100640/workflows/f4e73659-1f14-45da-8740-7314e788fb72/jobs/1340353 that happened in #32651 . Can't reproduce locally.
Caused by

transformers/examples/tensorflow/summarization/run_summarization.py

Lines 62 to 70 in 9d2ab88

	try:
	nltk.data.find("tokenizers/punkt")
	except (LookupError, OSError):
	if is_offline_mode():
	raise LookupError(
	"Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
	)
	with FileLock(".lock") as lock:
	nltk.download("punkt", quiet=True)

:

Will cause a failure at test collection

_______ ERROR collecting examples/tensorflow/test_tensorflow_examples.py _______
examples/tensorflow/test_tensorflow_examples.py:67: in <module>
    import run_summarization
examples/tensorflow/summarization/run_summarization.py:63: in <module>
    nltk.data.find("tokenizers/punkt")
/usr/local/lib/python3.10/site-packages/nltk/data.py:554: in find
    return find(modified_name, paths)
/usr/local/lib/python3.10/site-packages/nltk/data.py:541: in find
    return ZipFilePathPointer(p, zipentry)
/usr/local/lib/python3.10/site-packages/nltk/compat.py:36: in _decorator
    return init_func(*args, **kwargs)
/usr/local/lib/python3.10/site-packages/nltk/data.py:394: in __init__
    zipfile = OpenOnDemandZipFile(os.path.abspath(zipfile))
/usr/local/lib/python3.10/site-packages/nltk/compat.py:36: in _decorator
    return init_func(*args, **kwargs)
/usr/local/lib/python3.10/site-packages/nltk/data.py:943: in __init__
    zipfile.ZipFile.__init__(self, filename)
/usr/local/lib/python3.10/zipfile.py:1271: in __init__
    self._RealGetContents()
/usr/local/lib/python3.10/zipfile.py:1338: in _RealGetContents
    raise BadZipFile("File is not a zip file")
E   zipfile.BadZipFile: File is not a zip file
------------------------------- Captured stderr --------------------------------

Would be nice to have:

• isolate the behaviour of the bad file creation
• flag the flaky part of the failing test accordingly.

[gante]

See this convo: NLTK doesn't support parallelism :(

[gante]

^ I think that PR is a proper fix :D