grep/icase: avoid kwsset on literal non-ascii strings
When we detect the pattern is just a literal string, we avoid heavy
regex engine and use fast substring search implemented in kwsset.c.
But kws uses git-ctype which is locale-independent so it does not know
how to fold case properly outside ascii range. Let regcomp or pcre
take care of this case instead. Slower, but accurate.
Noticed-by: Plamen Totev <redacted>
Helped-by: René Scharfe <redacted>
Helped-by: Eric Sunshine <redacted>
Signed-off-by: Nguyễn Thái Ngọc Duy <redacted>
Signed-off-by: Junio C Hamano <redacted>