(doesn't leave a partial UTF-8 sequence at the end of the destination string).

This calls memcpy() for every char, which may not be performant, but I didn't
(yet) do any profiling. Hence, no premature optimization either.