…860)

The next step after utf8_length_from_utf16_with_replacement
is almost always going to be to allocate a UTF-8 buffer and
then convert the string. Sadly, we have to insert a third
pass, to_well_formed_utf16, which converts the unpaired
surrogates.

Since surrogates are relatively rare, and the _with_replacement
functions have already scanned the input, we could skip the
conversion if we were given this information along with the
utf-8 length.

In my measurements on Icelake this doesn't slow down
utf8_length_from_utf16_with_replacement at all.