The first change is to move to a PIMPL strategy so that
we can provide a more extensive unwinding API to other modules.
In particular, you can get access to a python frame iterator
and make repeated calls on it.  This is in preparation of the
sys._getframe implementation.

This also exposed an issue with our usage of libunwind -- basically
you have to do all of your unwinding in one function, but we were
doing it in several different functions and happening to get away
with it.  So, switch from a C++11 range-for loop which is nicer, to
a callback approach which is less nice in my opinion, but safer.