I think when using matplotlib in the kernel, we want the inline backend more often than not. This is especially true in the notebook, where saving figures in the output is part of the point. I'm a bit hazy on how to achieve this, but I think it's worth doing.

Prompted by jupyter/docker-notebook#7