* We need to keep a PythonContext (and its globals Dict) around while
we still have some pending callbacks happening. So now the external
code creates a PythonContext and then releases it when it's done, but
the context will hang around until the global redirector object is
destructed, which is responsible for deleting the context.
* The global redirector is deleted when a refcounting cycle is detected
and the dict is unreachable, which only happens after the context is
released.
* Any time a callback is passed to something and converted to a
std::function we add a reference on the global redirector to keep it
alive. When the callback has finished executing we remove the ref.
* This way, any pending callbacks that have been called but not finished
or converted (queued) and not called yet asynchronously will keep the
context object alive to be able to output, handle exceptions, etc.
* Additionally we need to detect when we're being called asynchronously
and handle exceptions separately instead of trying to propagate up the
call chain, because there might not be any more python code up the
chain (e.g. the render manager calling a python callback).
* Using a separate dict for globals/locals for each interpreter means we
still get separation of variables and no persistence where we don't
want it, but removing sub-interpreters means pyside can work as it
uses the PyGILState_ APIs which do not support sub-interpreters.
* We import everything up front then duplicate the __main__ each time we
create a new context so we keep the __main__ pristine and muck up an
individual copy.
* Because sys is now shared, the output redirectors that overwrite
sys.stdout and sys.stderr have a NULL context, and instead they look
up a specific global which contains the actual context pointer.
baldurk