Fixes #270

ordinal() builds its result as f"{value}{suffix}", so a translation can only ever supply a trailing suffix. As reported in #270, that makes languages whose ordinals place the number inside the word impossible to translate — e.g. Romanian al 2-lea (m) / a 2-a (f), where the number sits between a prefix and a suffix.

Changes proposed in this pull request:

• In ordinal(), if the (context-aware) translated string contains a %s/%d placeholder, format the number into it (suffix % value) so the translation controls where the number goes; otherwise keep the existing "<number><suffix>" behaviour.
• Add regression tests for both the interpolated path (Romanian-style al 2-lea / a 2-a, plus a %d variant) and the unchanged suffix path.

Backward compatibility:

• All 35 existing locales store bare suffixes ("nd" -> "e", "th" -> "º", …) with no placeholder, so they take the unchanged branch and their output is identical. The English default is unaffected (ordinal(2) == "2nd"). No msgid changes, so no existing translation needs re-merging.

Verified with a throwaway Romanian locale loaded through the real gettext path (activate("ro", path=...)):

>>> humanize.ordinal(2)                  # al %s-lea
'al 2-lea'
>>> humanize.ordinal(23)                 # al %s-lea
'al 23-lea'
>>> humanize.ordinal(2, gender="female") # a %s-a
'a 2-a'
>>> humanize.i18n.deactivate()
>>> humanize.ordinal(2)
'2nd'

Full local run: 786 passed, ruff/black clean.