Make the AverageLearner only return new points ...

and introduce learner.pending_points to make it more similar to the rest of Adaptive.

I've added a test that will fail with the current master.