This is the path we settled on in #24.
That is: any new toplevel window takes over the Cage display, hiding any
previous toplevels until it is closed. Only when the last toplevel is
closed, does Cage exit as well.
This allows us to check per-view whether is has dialogs open, instead of
diong it on a global basis as we are doing currently. This is necessary
for fully supporting multiple primary clients.
We shouldn't render a window before it is mapped (obviously), but we
render all windows in the view list. Hence, only insert the window once
it is mapped.
We could run into the case where a window is destroyed without being in
the window list, so we now track unmapping again and remove windows from
the list when they get unmapped.
This makes Cage much easier to maintain. Not only is it easier where to
look and to maintain a mental model of the code, there is also more
encapsulation, better abstractions and better extendability.
