Add LLM translation eval suite and page cost tooling

[Raghaddahi]

Relates to #446

Description

Adds tooling to evaluate and price LLM translation for the Wagtail user guide, supporting the model/provider decision in #446.

Why this approach:

• Standalone scripts instead of fork changes — the eval suite calls the real fork translate_text (same SYSTEM_PROMPT, HTML sanitisation and validation as the admin "Translate with AI" button) without writing anything to the DB, so it runs safely against production data.
• Scripts are now language-parameterized via EVAL_TARGET_LANGUAGE/EVAL_LANG_SLUG env vars, with outputs namespaced under out/<lang>/, so the same pipeline covers Arabic and French.
• scripts/page_cost.py is a read-only cost report straight from saved TranslationLog rows (no LLM calls); page selection is anchored on the max StringTranslation.updated_at to match a page's translation run.
• Generated eval outputs (out/, per-model cost CSVs) are gitignored; scorecards from the runs are kept in the report on this branch for the Automated translations via wagtail-localize #446 discussion.

Areas needing careful review:

• The LLM-as-judge (run_judge.py) pairwise scoring in one call (candidates A/B/C shuffled per segment) — judge reliability is the crux of the eval.
• The check_rules.py deterministic checks (HTML tag/id preservation, <b>/<i> kept English, truncation) — these feed the scorecard.
• ruff.toml gains per-file-ignores for scripts/*.py (T20 prints, E402 Django-shell imports) since these are intentional CLI patterns.

Testing

• poetry run ruff check scripts/ — all checks passed.
• poetry run ruff format --check scripts/ — clean after ruff format.
• Re-ran summarize.py against existing French candidate data inside the container; output matches the pre-refactor scorecard exactly (rank logic intact after lint cleanup).
• Full eval runs (collect → rules → judge → summarize) for both Arabic and French completed successfully against the current prompt; scorecards are in TRANSLATION_EVAL_REPORT.md on this branch.

QA left to do: full 392-segment eval once a model decision is needed; no unit tests for the scripts (they are run inside the Docker web container via manage.py shell).

AI usage

• Agent: Claude, via opencode
• AI generated the code and this PR description; human-reviewed by the contributor.