Demonstrate differentiating through model output

[joewallwork]

Towards #111.

We should further test #152 by calling tensor.backward on a tensor outputted from a network. By doing so, we'll check that we're able to back-propagate through a network from code that depends on it.