[skrifa] CJK autohinting weirdness

Some confusing bits in the FreeType CJK autohinter.

1. Scaled/fitted values are never generated for stem widths so they're always 0.
https://gitlab.freedesktop.org/freetype/freetype/-/blob/57617782464411201ce7bbc93b086c1b4d7d84a5/src/autofit/afcjk.c#L661

2. There's an extra loop to detect round segments that is never executed because the pointer to the segments array is captured (as NULL) before the segments array is actually allocated (by the call to `af_latin_hints_compute_segments`).
https://gitlab.freedesktop.org/freetype/freetype/-/blob/57617782464411201ce7bbc93b086c1b4d7d84a5/src/autofit/afcjk.c#L808

We'll just match this behavior for now but should investigate later to see what's going on.

Start by comparing the CJK autohinter logic in src/autofit/afcjk.c at the linked stem-width and round-segment sections. Trace how af_latin_hints_compute_segments allocates the segments array and determine whether the port should preserve the observed FreeType behavior or correct it. Done means the intended behavior is documented or covered by a focused test.

Written by the indexing model from the issue text.