resolve_kokoro_lang_code validates against mlx-audio's own LANG_CODES —
its docstring is explicit that the vendored table is the only source of
truth — but built the "Kokoro supports: …" list in the rejection message
from the hardcoded _KOKORO_ISO_BY_FULL_NAME instead.
That map exists to translate full names into Kokoro's codes, and it has
no entry that produces "b". British English is reachable as en-gb, and
the existing test asserts resolve("en-gb") == "b", yet the message never
mentioned it: the installed table lists nine languages and the message
named eight. Any language a later mlx-audio adds would be dropped the
same way, telling a user to switch engines when they need not.
Derives the labels from ALIASES and LANG_CODES instead, preferring the
full name a caller can actually pass and falling back to the table's own
description for codes no full name reaches.
The derivation takes the tables as arguments, so its tests run on every
platform. The resolution tests around it need the Apple-Silicon-only
package and skip on every CI runner, which is why a message drifting
from the table it documents went unnoticed.
Fixes#2156
