Added abuse guards to the i18n review workflow (#28209)

no ref
- Caps automated review at 15 files / 500 added lines per PR; oversized
PRs get a brief "skipped" notice instead of a model call
- Adds a 5-minute job timeout so a stuck Anthropic call can't eat runner
minutes
- Together this puts a hard ceiling on per-PR API spend without changing
the experience on normal translation PRs

The cap is deliberately scoped to translator-contributor PRs (typically
1–5 files). Developer PRs that introduce new keys and auto-propagate
them across all ~62 locales (typically 50–90 files) will be skipped —
those contain English placeholders rather than translations, so the bot
has little to validate there.