Utf8Proc::Analyze: check 32 bytes per iteration while the input is ASCII - #25721

The ASCII scan checked 8 bytes per iteration with a branch each. Folding
four loads into one mask check per 32 bytes is neutral on short strings
and about 9% faster when appending long ASCII strings.

The 32-byte scan cost every short string an extra loop test before the
original code, which showed up as +3% on TPC-H ingestion where almost all
strings are shorter than 32 bytes. Gate it behind a single length test so
short strings run exactly the previous code and only long inputs, where
the scan runs at least twice, take the wide path.

Locally: 40M 28-byte strings unchanged, 20M 208-byte strings -9%.