chars: reduce searching time with roughly 85 percent for plain ASCII

Make case-insensitive searching in a UTF-8 locale eight times faster
when the actual characters involved are plain ASCII.

This makes us faster than 'less', and as fast as Vim and Emacs.

The disadvantage of this change is that searching for a string that
begins with a multibyte character is nearly ten times slower than
searching for one that begins with an ASCII character.  This may be
unsettling when searching a huge file first for a simple ASCII string
and later for a UTF-8 one.  Doing this second search, the user might
get impatient: "Why is it taking so long?"

(This patch fell through the cracks four years ago, when I worked on
the searching code.  It sat in a branch on top of other changes that
I never applied because I made different improvements.  The speedup
at the time, on that machine, was only around sixty percent, though.
But measuring it now again on the same machine, it clocks in at an
82 percent reduction with -O0 and an 87 percent reduction with -O2.)

src/chars.c

@@ -374,6 +374,22 @@ int mbstrncasecmp(const char *s1, const char *s2, size_t n)
 		wchar_t wc1, wc2;
 
 		while (*s1 != '\0' && *s2 != '\0' && n > 0) {
+			if ((signed char)*s1 >= 0 && (signed char)*s2 >= 0) {
+				if ('A' <= (*s1 & 0x5F) && (*s1 & 0x5F) <= 'Z') {
+					if ('A' <= (*s2 & 0x5F) && (*s2 & 0x5F) <= 'Z') {
+						if ((*s1 & 0x5F) != (*s2 & 0x5F))
+							return ((*s1 & 0x5F) - (*s2 & 0x5F));
+					} else
+						return ((*s1 | 0x20) - *s2);
+				} else if ('A' <= (*s2 & 0x5F) && (*s2 & 0x5F) <= 'Z')
+					return (*s1 - (*s2 | 0x20));
+				else if (*s1 != *s2)
+					return (*s1 - *s2);
+
+				s1++; s2++; n--;
+				continue;
+			}
+
 			bool bad1 = (mbtowc(&wc1, s1, MAXCHARLEN) < 0);
 			bool bad2 = (mbtowc(&wc2, s2, MAXCHARLEN) < 0);