tabulate correctness tests fail when wcwidth is installed #228

What

Four comparisons in tabulate/test_tabulate_correctness.py fail when wcwidth is installed in the environment:

FAILED TestEdgeCases::test_unicode_basic
FAILED TestEdgeCases::test_cjk_characters
FAILED TestEdgeCases::test_mixed_cjk_ascii
FAILED TestEdgeCases::test_wide_emoji

Why

The reference tabulate computes column widths with wcwidth when that package is importable and with len(s) otherwise. Our module's default is len(s) (WIDE_CHARS_MODE is opt-in), and the test module's guarded import only shields the reference from the repo's own tabulate/ directory, not from wcwidth.

The two sides therefore apply different rules, and the tables differ:

café (e + combining acute)
  ours  "café   1"      len() = 5
  ref   "café  1"       wcwidth = 4, the combining mark is zero-width

你好
  ours  width 2         len()
  ref   width 4         wcwidth, East Asian wide = 2

Measured: wcwidth present (Python 3.10 env, pulled in as a transitive dependency) → 4 failed, 79 passed; wcwidth absent (3.12 env) → 83 passed. CI is green because the CI environment does not install wcwidth.

Fix

Hide wcwidth for the reference import — sys.modules["wcwidth"] = None makes the import fail cleanly, so the reference selects its len(s) path and both sides compute widths the same way in every environment — and add a guard test asserting tabulate.wcwidth is None, so the comparison cannot silently drift back.

This follows the same pattern as the other reference-shadowing fixes (#196 soupsieve 2.10, #205 benchmark references).