The new implementation is copied from another project and is
BSD-licensed.  It is exception-safe and can parse IPv6 scope ids with
interface names.

The API documentation says "move as much data as possible", and if
there is room at the head of the buffer, we should use that if the
room after the tail is not large enough.