resolve segfault

commit 1268b4b4 just shifted the
problem. DNG nicely shut down but a unit test segfaulted. I did not
investigate in depth what exactly happened. However, the cleanup
at Python shutdown seems to be necessary. The problem is that the
scene destructor might have already been called when the cleanup at
Python exit happens and the pointers to the scene observers point into
Nirvana. So by explicitely defining the Scene destructor that kills
all observer pointers and bound gfx objects, the likelihood for bad things
happening when the RemoveAll function is called on the Singleton again
is reduced.