101 no regularization grad

[wsdewitt]

Description

Make model.FullyConnected.regularization_loss() differentiable.

Closes #101

Tests

tests/test_model.py

Checklist:

• The code uses informative and accurate variable and function names
• The functionality is factored out into functions and methods with logical interfaces
• Comments are up to date, document intent, and there are no commented-out code blocks
• Commenting and/or documentation is sufficient for others to be able to understand intent and implementation
• TODOs have been eliminated from the code
• The corresponding issue number (e.g. #278) has been searched for in the code to find relevant notes
• Documentation has been redeployed

[wsdewitt]

My current challenge is that loss.backward() returns None in my new test module, so I'm having trouble validating the gradient.

[wsdewitt]

Ok I think this is ready for review

[zorian15]

Sick, looks good -- I don't have any complaints.