Seemingly unrelated regressions enhancement

This merge request implements seemingly unrelated regressions (SUR) and a discovery procedure that uses SUR rather than the residual iterative conditional fitting (RICF) algorithm. The large diff mostly results from output of the various tests that I performed including confidence interval tests to compare SUR and RICF, tests to compare the discovery performance of SUR and RICF, tests that optimize across one or two lambda values in SUR (we found that using two produced the best discovery results), and tests that validate the chain graph constraint. The former three test types reside in sur_tests.py while the latter type resides in chain_constraint_tests.py. It may be logical to split each of these test types into their own files at some point. The SUR procedure is found in sur.py while the SUR discovery procedure is found in admg_discovery_sur.py.