Fix outdated examples, broken links, and stale references across docs

[qiyanjun]

Summary

Reviewed docs/0_get_started/, docs/1start/, and docs/3recipes/ for outdated code examples, broken external links, and stale version/count references. Every item below was verified against current code or a live HTTP request before fixing — none of these are guesses.

What was found and fixed

• installation.md: "Python 3.6 or above" and the conda example's python=3.8 are both stale — setup.py has required >=3.9 since the black-version-upgrade PR (chore: pin black/isort/flake8 to modern versions #833). Bumped both to 3.9.
• command_line_usage.md: textattack eval-model was renamed to textattack eval (confirmed against EvalModelCommand.register_subcommand's actual registered subparser name — it's literally "eval", not "eval-model").
• quick_api_tour.rst: fixed two Sphinx cross-references pointing at a nonexistent transformers.models.wrapper module (should be textattack.models.wrappers), and a malformed :ref:TextFooler attack role with no matching target.
• what_is_an_adversarial_attack.md: a copy-pasteable CLI example had its --flag hyphens autocorrected to em-dashes at some point (attack — model lstm-mr — num-examples...), making it non-functional if pasted as-is. Restored --.
• support.md: pointed contributors at transformation.rst for adding new transformation docs — that file doesn't exist; it's docs/api/transformations.rst.
• FAQ.md / attacks4Components.md / attack_recipes_cmd.md: "16 adversarial attacks" is stale — confirmed 21 CLI-registered recipes via ATTACK_RECIPE_NAMES + 3 Python-only multi-lingual recipes (French/Spanish/Chinese) = 24. Added the missing a2t, leap, and multi-lingual recipe rows to both HTML comparison tables, which listed neither despite both existing in the codebase already.
• attack_recipes.rst: fixed a stale recipe-numbering gap (17-19 should've been 18-20 after an earlier recipe was inserted above without renumbering everything below it); the list is now sequential 0-23.
• attack_recipes_cmd.md: the DeepWordBug example labeled "on DistilBERT trained on the Quora Question Pairs paraphrase identification dataset" actually referenced distilbert-base-uncased-cola — a grammatical acceptability model, unrelated to QQP/paraphrase identification. Fixed to distilbert-base-cased-qqp (confirmed this model exists on the Hub).
• models.md: github.com/huggingface/nlp is HuggingFace's old, renamed repo (confirmed it 301-redirects to github.com/huggingface/datasets); updated both the link text and href.

Verification

• Every factual claim above checked against either the actual codebase (ATTACK_RECIPE_NAMES, EvalModelCommand's registered name, setup.py's python_requires) or a live HTTP request (HF model pages, GitHub repo redirect)
• Full clean Sphinx rebuild afterward introduces no new warnings in any of these files — cross-checked against the pre-existing warning list from Fix all Sphinx build warnings/errors (19 -> 0) #837 (which touches an entirely different set of files) to confirm no overlap

🤖 Generated with Claude Code