Removes the assert(3) failure reported on the list for 1.7-RC1 when a
non-space, non-cntrl, rune has a wcwidth(3) of -1; output L'?' in its
place. The old code also didn't reset mbtowc(3)'s state before trying
to parse "?"; that's handled differently now so isn't an issue. Pad
with the multi-byte encoding of L" ", having ensured the wcwidth(1) of
L' ' is one, rather than with a non-wchar_t ' '. Point out padding only
occurs in one particular case, and not the other two; quite odd.