This change makes the test pass without the unicode feature.
Note that it is not possible to run the test without the unicode feature
unless one changes the Cargo.toml, so I did that locally to run the
tests
This is needed in the future to be able to roll back state within the
line breaking when we can't fit even a single word into the line and
have to go back to fitting glyph by glyph.
For the unicode line break iterator the clone is a little more complicated.
The iterator is clone, but the type is anonymous. I first had an
implementation to moved the iterator into a closure, which was cloned
via another control closure. But that's complicated and still involves
allocating the various bits & pieces the closure captures.
So instead the simpler solution used here is to allocate the breaks
into a shared vector.
