intword: name every short-scale power up to googol (fixes the decillion gap)

[lethevinh]

Summary

Fixes the still-open part of python-humanize#356: every value between 10**36 and a googol used to render as a giant decillion count (intword(10**50) → '100000000000000000.0 decillion'), and a value just below a googol never carried to '1.0 googol'.

Upstream PR python-humanize#346 already fixed the carry between adjacent named powers; the decillion→googol naming gap itself stayed open (the issue remains open upstream, with PRs python-humanize#364/python-humanize#400/python-humanize#401 closed unmerged).

What changed

• powers/human_powers now carry the standard short-scale names for 10**36–10**99 (undecillion … duotrigintillion), so every magnitude has a unit and the mantissa never exceeds 999.x.
• The existing integer-ratio carry now also fires across the final 10**99 → 10**100 step: intword(10**100 - 10**93) → '1.0 googol'.
• Tests: test_intword_powers expectations updated for the now-named range and extended with 10**50, 10**99, near-googol carry; test_intword_rounding_rollover mantissa guard widened to the full range.

New NS_ msgids fall back to English in untranslated locales (standard gettext behaviour).

Verification (repo's own suite)

python -m pytest → 749 passed, 110 skipped (baseline at pin 392aef7: 746 passed) — Python 3.12, offline.

Pre-existing limitation observed but out of scope: intword(10**400) raises OverflowError (float() conversion) at the pin as well — unchanged by this diff.

This PR targets the osalks fork deliberately: the adopting workspace (osalk-company CS-B1) records fork-level evidence first; upstream submission is deferred to a separate human decision. No upstream tracker traffic was generated.