unstructured2graph: NLTK data (averaged_perceptron_tagger_eng / punkt_tab) not available in CI #257

What's wrong

test-unstructured2graph fails 22 tests with:

LookupError:
**********************************************************************
  Resource 'averaged_perceptron_tagger_eng' not found.
  Please use the NLTK Downloader to obtain the resource:

  >>> import nltk
  >>> nltk.download('averaged_perceptron_tagger_eng')

  Attempted to load 'taggers/averaged_perceptron_tagger_eng/'

  Searched in:
    - '/home/runner/nltk_data'
    - '/home/runner/work/ai-toolkit/ai-toolkit/unstructured2graph/.venv/nltk_data'
    - '/home/runner/work/ai-toolkit/ai-toolkit/unstructured2graph/.venv/share/nltk_data'
    - '/home/runner/work/ai-toolkit/ai-toolkit/unstructured2graph/.venv/lib/nltk_data'
    - '/usr/share/nltk_data'
    - '/usr/local/share/nltk_data'
    - '/usr/lib/nltk_data'
    - '/usr/local/lib/nltk_data'
**********************************************************************

Example failure (test_from_texts_only_chunks_creates_real_chunk_nodes), full call chain:

tests/test_e2e.py: from_texts(["Alice works on the graph engine.", ...], memgraph, only_chunks=True)
  -> src/unstructured2graph/loaders.py:309 (from_texts) -> parse_text(text)
  -> src/unstructured2graph/loaders.py:99 -> unstructured.partition.text.partition_text
  -> unstructured/partition/text.py -> element_from_text -> is_possible_narrative_text
  -> unstructured/partition/text_type.py:186 (contains_verb) -> nltk pos_tag
  -> nltk/tag/__init__.py -> PerceptronTagger() -> nltk/data.py:find() -> LookupError

unstructured.nlp.tokenize actually has a built-in safety net for exactly this
(.venv/lib/python3.10/site-packages/unstructured/nlp/tokenize.py):

# auto download nltk packages if the environment variable is set
if os.getenv("AUTO_DOWNLOAD_NLTK", "True").lower() == "true":
    download_nltk_packages()

download_nltk_packages() checks for averaged_perceptron_tagger_eng (taggers) and
punkt_tab (tokenizers) and calls nltk.download(..., quiet=True) for whichever is
missing -- this runs automatically at import time and defaults to on. It's evidently
not succeeding in the CI runner: quiet=True suppresses nltk.download()'s own
diagnostic output, so a failed download (network/CDN blip, egress restriction to
NLTK's data host, timeout, etc.) fails silently and this LookupError is the first
visible symptom, later, wherever pos_tag/contains_verb actually gets called. This
would also explain the "52 passed, 22 failed" split in one run: only the tests whose
code path reaches pos_tag fail, and since the download attempt only happens once per
process (module-level, first import), every subsequent call in that same test run
keeps failing once it's failed once.

Not something introduced by #256 -- that PR only touched unstructured2graph's
README/example, no dependency or code changes. Unclear whether this is newly broken
(e.g. a CDN/host change, or an interaction with the nltk<3.10 pin from #251) or a
latent gap that just hadn't been hit before; worth checking both.

Acceptance criteria

• Confirm why download_nltk_packages()'s auto-download isn't succeeding in the
  test-unstructured2graph CI job (network egress? rate limiting? interaction
  with the nltk<3.10 pin?)
• Make CI not depend on a silent, unverified runtime auto-download -- e.g. an
  explicit python -m nltk.downloader averaged_perceptron_tagger_eng punkt_tab
  step in .github/workflows/tests.yaml's test-unstructured2graph job (with
  visible output, not quiet=True), so a real failure is loud instead of
  surfacing as 22 unrelated-looking test failures
• test-unstructured2graph's full suite passes cleanly in CI

Related

Found while investigating CI failures on #256 (unrelated to that PR's own changes).