[SAY-5]

Fixes #181

Changes proposed in this pull request:

• Let translation catalogs define their own named powers of ten and pluralized number/unit patterns. Match each pattern table to its exact scale vector so an incomplete regional catalog cannot silently inherit labels for different powers.
• Use the Japanese four-digit scale, so 234909023 becomes 2.3億 and 2349090 becomes 234.9万. Preserve the existing formatting path for catalogs that do not opt in.
• Document the catalog entries and test real compiled translations, rounding, negative values, regional fallbacks, malformed profiles, locale isolation, and the translation update pipeline.

Validation: 873 ordinary tests passed on Python 3.13; Ruff 0.16.5, Black 26.5.1, Prettier 3.9.6, and mypy passed. The 15 benchmark tests were not run because their plugins were unavailable. A separate bounded timing check measured about 0.16 microseconds of additional overhead for English intword calls.

This uses new opt-in catalog entries and leaves existing unit message IDs in place, keeping the change separate from the placeholder migration in #403. The existing French plural-selection rule and the English decillion/googol gap are unchanged.

[Wkhanex]

With a small injected catalog scale/pattern table, intword selects the correct named power, preserves the negative sign and leaves values below the first scale as numbers. Empty, unsorted, duplicate, boolean, excessive and malformed scale metadata all fall back to the legacy formatter. These checks use catalog-function substitutes rather than compiled Japanese catalogs.

Checked 5eb2a8de211b with 11 standalone assertions on Python 3.12.14. I haven't run the full suite. The generated _version module was supplied as metadata only.