WIP: TVD objective in GST

[rileyjmurray]

I noticed there's a total variance distance objective function class which is almost finished. The only missing functions for the code to run without formally raising an error were RawTVDFunction.dterms and RawTVDFunction.chi2k_distributed_qty. The former function has a very simple "canonical" implementation, but it's numerically problematic since the 1-norm isn't differentiable at any vector with a zero component. I don't know what a mathematically correct version of the second function looks like (or if there is such an implementation) so I just have it return -1 to indicate "this doesn't mean what you think it means."

I'm mostly opening this PR to facilitate QPL-internal discussion on the topic.

You can run GST where the final iteration is TVD-based instead of log-likelihood using something like the following.

models, opt_results, cache = pygsti.algorithms.run_iterative_gst(
    dataset, starting_model, lsgst_circuit_list_of_lists,
    optimizer={'tol': 1e-5}, resource_alloc=None, verbosity=4,
    iteration_objfn_builders=['chi2'], final_objfn_builders=['tvd']
)