i18n: the parity numbers are written by hand and drift; here is the matrix measured from the catalogs #2035

docs/localization-parity.md carries hand-written counts, and they drift: it said the Japanese
panel catalog had 397 rows when it had 506.

Proposal: a cargo xtask i18n-coverage that prints what the catalogs actually hold, so the doc can
quote a number that comes from the source instead of a person. The four surfaces are easy to read:
i18n.rs and i18n/ui.rs are tuple tables, i18n/effects.rs is keyed by effect id, and the Settings
strings are the s(...), Section(...) and choice literals in engine/src/prefs.rs that the panel
catalog has to carry.

Two caveats, because such a table is easy to over-read:

• A row is not a translation. Cryptomatte stays Cryptomatte in every catalog on purpose, and a
  brand, a unit or a number is allowed to.
• It counts rows, not coverage of what is drawn. A tr() call site with no row falls back to
  English silently, which is the class of gap i18n: the labels the call-shape scan cannot see #1581 and i18n: two labels that already had a catalog row were formatted in English #1669 are about.

The two gaps the matrix showed are filed separately: Traditional Chinese's missing effect rows are
#2037, and German/French/Russian having no rows below the menu is #2038.