Ges

- Fixed bug: The calls ges.fit and ges.fit_bic now return the total score, instead of the total score change wrt. the empty graph.
- Set a minimum value (float machine epsilon) for the estimated noise term variances. For very low sample sizes noise-term variances can be estimated as zero because of machine precision, breaking the logarithm in the likelihood computation.
- When checking if an operator improved the score, require it to be greater than a threshold (instead of zero); because of numerical issues, two opposing turn operators (idemptotent when applied together) may both return a score slightly over-zero; as a result the algorithm gets stuck in a loop.

- Updated MANIFEST.in file to include readme and license.

- ges.fit now allows for an optional parameter `completion_algorithm` which allows for a user-defined algorithm to go from PDAG to CPDAG after the application of each operator.
- removed deprecated use of np.bool

- Both ges.fit and ges.fit_bic now allow for an optional parameter `iterate`, to indicate whether the given phases should be iterated more than once. Defaults to `False` to maintain backwards compatibility.
- Adapted tests.

- Fixed bug in "raw" computation of Gaussian observational score + expanded tests
- Ran autopep on all files, removed unused imports, etc.
- Expanded docstrings and doctests.
- Changed ges.py to main.py to avoid module/package name conflict.
v1.0.2, 2021-02-05 -- Adapted testing code to new sempler version 0.2.0.
v1.0.1, 2021-01-24 -- Fixed project url and license on setup.py, added changelog.
v1.0.0, 2021-01-24 -- Initial release.

-- Can pass empty dict as {shift,do}_interventions parameter and will have the same effect as not passing anything or passing None
-- Refactored matrix_block function
-- Fixed license and project URL in setup.py
-- Set a markdown readme for PyPi
-- Major changes in the readme
-- New sphinx documentation and doctests
-- Updated imports in setup.py, made dependency requirements more specific
-- The init of NormalDistribution can be asked to check that the covariance matrix is positive definite.
-- The regression method in NormalDistribution now uses np.linalg.solve to find the coefficients
-- ANM.sample, LGANM.sample now accept noise_interventions.
-- New module sempler.generators
-- New sphinx documentation and doctests
-- API changes:
-- sempler.dag_avg_dev and dag_full have been moved to sempler.generators
-- ANM.sample now requires the sample size to be specified
-- The methods in NormalDistribution throw new exceptions
-- NormalDistribution.equal now has optional arguments rtol and atol instead of tol.
-- The LGANM constructor is now LGANM(W, means, variances) instead of LGANM(W, variances, means)
v0.1.3, 2020-09-28 -- Fixed issues 1,2,3. Arrows now appear when plotting acyclic graphs.