This fixes an edge case (20206) where if the input ended in a character
needing encoding but there was no newline on the string, the last byte
of the encoded character would be dropped.  The fix is to use a more
efficient algorithm, provided by Serhiy Storchaka (5803), that does not
have the bug.