MO_SS_Conv: Don't give up the highest bit for sign

According to this test, reducing the value for the sign is not correct.

narrow[W32→W8](sext[W16→W32](load[W16](0x223972::W64)))
test ( bits64 buffer ) {
    bits64 ret;
    (ret) = prim %popcnt8(%lobits8(%sx32(bits16[buffer + (2242930 :: bits64)])));
    return (ret);
}

4 /= 5