Fix PyUnicode_Substring() for start >= length and start > end

Remove the fast-path for 1-character string: unicode_fromascii() and
_PyUnicode_FromUCS*() now have their own fast-path for 1-character strings.