This PR contains a few commits as part of an attempt to make full use of pytest to run the tests. It increases the coverage in test_ampgo.py, test_basinhopping.py, test_brute.py and test_itercb.py.

As an outcome of the increased coverage a few oversights were found (incidentally?) all introduced by me ;) Anyway, the following fixes were made:

1. logic in show_candidates is now explicit
2. the estimation of the covariance matrix is only attempted when the fit finished correctly and not when it was aborted.

The code would probably benefit from some re-factoring, but that should be done at the same time as we address the _calculate_statistics function. For example, right now result.residual is set when the fit is aborted, but still needs to be calculated if it finishes normally. In many instances we do a check like if not result.aborted and then calculate/fill in the MinimizerResult attributes. After that we calculate the statistics, which should be done regardless, and then -only if not aborted- we should calculate the covariance matrix. There are a bit too many if-statements involved there, but I cannot get easily a consistent way of doing this for all solvers. So my changes for now work as intended, and making big changes just before a release does not seem smart... Ideally, we would tackle this once the test coverage has increased and we are not as likely to miss potential issues.

Finally, I updated the dependencies in .travis.yml to add numdifftools and install emcee only for version == latest. These dependencies and pytest are installed with pip because conda does not have consistent versions available for all Python versions we test.