State count selection using information criterion

[dwysocki]

After completion of #1, I will implement automatic selection of the number of states using an information criterion, such as AIC or BIC.

There will have to be a new option to the train subcommand, --state-range LOW HIGH, where LOW and HIGH are the bounds. This must be used in conjunction with --init MODE (see #1), and cannot be used in conjunction with --states N, as it chooses N.

There will also be a --criterion CRITERION, where CRITERION may be one of AIC or BIC, assuming I decide to implement both.

I wish to use gradient descent to optimize with respect to the criterion, but am not sure if there will be multiple extrema, in which case it would not work like I desire. To test this, I need to be able to iteratively increase the number of states, and plot n_states vs AIC/BIC, to visualize whether multiple extrema exist. This could be implemented as an option like --output-criterion FILE, and write tabular data to FILE. Even after determining whether gradient descent works, this could be useful for my report.

[dwysocki]

For the time being, I have implemented an algorithm which subdivides the search space, finding the lowest number of states which maximize the likelihood function. This is not as robust as using something AIC, but will do for now. I am saving AIC/BIC for the journal paper.