Integration test for simple example

[willGraham01]

Resolves #38, though #75 remains something that we need to address.

Implements a barebones integration test for the simple example problem involving two normals.

All the steps to setup the causal estimand and constraints are done manually, IE by manipulating the Graph.model rather than attempting to use the CausalProblem class. Iterative PRs on top of this one will slowly take more of the manual steps out of the new integration test and into the appropriate classes.

Also, the integration test currently uses the "minimise" approach described in #75, whereby to find stationary points of the Lagrangian we elect to instead use SGD to minimise $\vert\nabla\mathcal{L}\vert^2$. This works for this particular example, but we also have the benefit of knowing the exact solution so are able to provide a very good initial value.

[willGraham01]

#70 changes the names of the nodes + parameters inside the two_normal_graph_parametrized_mean fixture, will need to update text in light of this PR if it goes in first.