Test a 'real-world' use case

[f-dangel]

Add a test that takes a neural network, converts its layers and sets requires_grad=True for only a subset of parameters. Then compute a forward pass and backward pass, make sure the gradients are correct.