Add eva main function and add eva.fit_gev fitstart options

[stellema]

• Added main() function to eva.py so that GEV parameters can be saved in a workflow.
  • Added eva.py to setup.py entry_points, so unseen will have to be reinstalled before it will work (e.g., run pip install --no-deps -e .)
• Added lmoments3 to ci/environment.yml
• Updates to eva.fit_gev:
  • Renamed keywords to more descriptive names:
    • test_fit_goodness => assert_good_fit
    • test_relative_fit => pick_best_model
  • Added keyword fitstart that replaces generate_estimates and user_estimates. The new keyword can be a string or a list of initial estimates.
  • Added a bunch of options that utilise the fit or fit start methods from scipy and xclim and one that uses the L-moments method. I've found the linear moments estimate LMM is very good, but kept scipy_fitstart as the default because it reproduces the original scipy.stats.genextreme.fit method. The scipy_fitstart and scipy difference is a bit confusing because using 'scipy' will basically fit the parameters twice.
  • The functionality of generate_estimates can be reproduced using scipy_subset or retry_fit.
  • There is now a specific option retry_fit, which retries the fit using a different initial estimate - generated by running the same fitstart method with data[::2]. If the retry attempt is not better than the first estimate (based on the log-likelihood), then the original parameters are returned. We could retry the fit using different fitstart methods, but I don't think this is a good idea because it could impact reproducibility.
• The eva tests have been updated for consistency and now utilise pytest fixtures.
• Updated fit_gev output so that is more consistent with xclim.indices.stats.fit, including renaming theta => dparams and adding attributes tofit_gev output.
• Updated check_gev_fit: The goodness of fit test is now based on the one-sample Kolmogorov-Smirnov or Cramér-von Mises test.
• Deleted fit_stationary_gev
• Added gev_confidence_interval - it is a vectorised version of the CI bootstrapping methods used in the eva plots (return periods or return levels). It still needs some work - I can finish it if it is needed.

[stellema]

I also added a copy of the ari_to_aep and aep_to_ari functions from https://github.com/climate-innovation-hub/frequency-analysis/blob/master/eva.py.