Add iterative approach to the inside_outside method

[hyanwong]

With variational gamma / EP, we iteratively pass messages until convergence. This could equally be done for the "inside-outside" (discrete time bins) method. This would solve the issue of Loopy Belief Propagation (see #165)

Correct me if I'm wrong here @nspope, but the main implementation difference between iteratively passing and just doing a single up/down pass of the ARG is that if we are just doing a single pass, it is better to raise the edge contribution to the power of the edge span. In contrast, when iterating, raising to the power of the span seems to prevent convergence.

If this is the case, then API-wise we pass a control dictionary, we could allow a "number-of-iterations" parameter which is 1 by default in inside_outside, but something else for variational_gamma. When set to one, we use the power-of-span method, but when >1 we don't.