1) Restrict the dynamic dynesty test to maxcalls of 100
2) Split the sampler_test out of the main tests and into a separate
node. This means the sampler_test doesn't contribute to the coverage

`python setup.py install` is not exactly the same, `pip install` validates more of the metadata

This adds a test-mode command like argument to run with zero likelihood.
As such, this will sample from the priors alone. To use, run any
standard bilby example with `--bilby-zero-likelihood-mode` in the
command line. This has been intentionally obfuscated to avoid
overloading the usual documentation.

- Make the test case generate samples from the prior
- Force the marginalized likelihood test to use fixed m1 and m2 to
  ensure no samples are generated with the wrong ordering
- Remove pipenv

to make sure the core dependencies are specified properly

1) Rename all tests to the pytest standard naming convention
2) Update to use pytest to run the tests and collect coverage
3) Simplify gitlab CI script using above changes

Partial solution to #181

1) If sigma is not given, then it will not be sampled over (or a warning
raised that you didn't specify a prior for sigma).
2) Comments added to explain the logic throughout
3) Tests added for the Gaussian likelihood

This updates to an anaconda python image and removes redundant tests
rerunning examples so that the tests finish in a more reasonable amount
of time. Many requirements are also moves to an
optional_requirements.txt file.