This PR does two somewhat controversial things:

1.  It adds in a new "rclpy.managed_init" method to rclpy,
which would be used with a context manager instead of
rclpy.init.  I had to do things this way because there
is no way (as far as I know) for the callee to know whether
it was called within a context manager or not, and I
don't think we can reasonably change the semantics of
rclpy.init() at this point.

2.  It changes the context manager implementation of
Context so that it does *not* call init() within itself.
This definitely changes the semantics, but as it stands
that initialization doesn't make sense because it can't
take arguments.  I think this change is warranted,
though we may have to search through documentation and
examples to make sure this doesn't break anything.

Signed-off-by: Chris Lalancette <clalancette@gmail.com>

That way, we don't need managed_init() anymore.

Signed-off-by: Chris Lalancette <clalancette@gmail.com>

That way we can use context managers, but actually properly
clean up, including uninstalling signal handlers.

We also switch to tracking node resources in the context,
so that when the context goes away, all nodes associated with
it are automatically destroyed.

Signed-off-by: Chris Lalancette <clalancette@gmail.com>

Signed-off-by: Chris Lalancette <clalancette@gmail.com>

Co-authored-by: Shane Loretz <sloretz@openrobotics.org>
Signed-off-by: Chris Lalancette <clalancette@gmail.com>

Co-authored-by: Shane Loretz <sloretz@openrobotics.org>
Signed-off-by: Chris Lalancette <clalancette@gmail.com>

Signed-off-by: Chris Lalancette <clalancette@gmail.com>