Thoughts about user interface

[mscroggs]

Creating this issue to dump thoughts about eventual user interface to come back to later:

• Creating the graphs could be completely hidden from most users: when the user asks for (eg) the expectation of X, we could construct the graph of X and its dependencies in the background at that point, then use it in the computation. Each node will know how it depends on its dependencies so we can iterate recursively through parents and work out the entire graph.
  This is similar to what is done in UFL: when a UFL form is created, the graph is created in the background, but the user does not explicitly have to ask for it
• Should we require that users give every node a label, or automatically label nodes that aren't given labels? Make two node example graph #16 originally did that later, but I'm now thinking that this may be a premature optimisation so am enforcing the giving of names for now

See also #13

[mscroggs]

Creating the graphs could be completely hidden from most users: when the user asks for (eg) the expectation of X, we could construct the graph of X and its dependencies in the background at that point, then use it in the computation.

We could aim for the user interface for the two node test to be something like this:

ux = DistributionNode(NormalFamily(), "UX", constant_parameters={"mean": mean, "cov": stdev**2})
x = DistributionNode(NormalFamily(), "X", parameters={"mean": ux}, constant_parameters={"cov": stdev2**2})
result = causalprog.algorithms.expectation(x, samples=samples, rng_key=rng_key)

or even:

ux = NormalDistribution("UX", mean, stdev)
x = NormalDistribution("X", ux, stdev2)
result = causalprog.algorithms.expectation(x, samples=samples, rng_key=rng_key)

All the information needed to create the tree is known by x, so the tree could be created inside ausalprog.algorithms.expectation.