Improve output of marginals when we have too few samples/chains to compute diagnostics

[null-a]

Once #25 is merged marginals will include n_eff and r_hat columns. In some (not very common) situations we don't have enough chains and/or samples to compute these, in which case we show nan in the table. I'm not sure this is very informative, so perhaps it would be better to show something like n/a instead.

This isn't a trivial change because the current assumption is that the table holds only floats and so is backed by a numpy array. Perhaps this could to be switched to a pandas data frame.