Following on to my patch of 9/7 implementing reverse-step, reverse-next
and reverse-continue, this patch is sufficient to implement
reverse-finish.  This is meant for discussion, not approval (I don't
plan to check anything in right away).

Again, this is based on a modular approach that assumes only a
simple interface "get_exec_direction()".  I haven't taken it
down to the architecture level, but have simply assumed that
every function has a single entry point (at its label), and
that it is sufficient to continue backwards to that point,
and (assuming the correct frame) do one more single-step to
reach the call.

Here's the patch: