mbstring's Unicode case conversion is table-driven, using Minimal Perfect Hash tables.
However, for small codepoint values, we bypass the hashtable lookup and just use
hard-coded conversion logic (i.e. adding or subtracting 0x20 from the appropriate
ASCII range).

For upcasing and downcasing, we had already optimized the conditional which sends
execution down this fast path, to use the fast path for as many codepoint values
as possible. However, for case folding, this had not been done.

This will give a small performance boost for case-folding Unicode text which
includes non-breaking spaces, symbols like ¥ or ™, or accented Latin
characters (used in many European languages).