chars: probe for a valid UTF-8 starter byte, instead of overstepping

Instead of always stepping back four bytes and then tentatively
moving forward again (which is wasteful when most codes are just
one or two bytes long), inspect the preceding bytes one by one
and begin the move forward at the first valid starter byte.

This reduces the backwards searching time by close to 40 percent.

src/chars.c

@@ -383,8 +383,20 @@ size_t move_mbleft(const char *buf, size_t pos)
      * possible point. */
     if (pos < 4)
 	before = 0;
-    else
-	before = pos - 4;
+    else {
+	const char *ptr = buf + pos;
+
+       if ((signed char)*(--ptr) > -65)
+	    before = pos - 1;
+       else if ((signed char)*(--ptr) > -65)
+	    before = pos - 2;
+       else if ((signed char)*(--ptr) > -65)
+	    before = pos - 3;
+       else if ((signed char)*(--ptr) > -65)
+	    before = pos - 4;
+	else
+	    before = pos - 1;
+    }
 
     while (before < pos) {
 	char_len = parse_mbchar(buf + before, NULL, NULL);