WeightNormalization fails after update

[seanpmorgan]

Yesterday I cut a 0.3 release on pypi test, but as a sanity check on the weightnorm changes #208 I ran a quick regression check:

https://colab.research.google.com/drive/1RQlVR9X7vj8q3W3sJC6YA3Cq45XTXrau

The new layer is failing with loss of nan. Looking at the test coverage for this layer I'm struggling how to best catch this without running a full model. The losses for our toy test cases do not return nan so perhaps we need a true regression test like mnist?

EDIT -- Fix colab link