Support for multi-level cluster survey designs

[AngusMcLure]

This could be done relatively straightforwardly for the logitnormal (and cloglognormal...) case by using standard random-effects style additive normal random effects for each level.

For two levels there would be just one additional unknown (with the $\sigma$ replaced with $\sigma_1$ and $\sigma_2$ for the two levels of clustering$). There would be only one additional set of integrals (compared to the existing two) however, they would be 2-d integrals.

Could use something like cubature to do the multidimensional integration efficiently and accurately without resorting to monte carlo.