Add BernoulliLogit

[devmotion]

This PR adds a Bernoulli distribution with the logit of the success rate as parameter.

This is motivated by use cases e.g. in Turing (TuringLang/Turing.jl#1890) where it is more convenient to work with unconstrained parameters. First class support for the parameterization allows us to evaluate e.g. the log pdf more efficiently.

[mschauer]

Can you mention “parametrized by the log of the odds Logitp = p/(1-p)” in the docstring

[codecov-commenter]

Codecov Report

Base: 85.94% // Head: 85.50% // Decreases project coverage by -0.44% ⚠️

Coverage data is based on head (9ed9c73) compared to base (4c27f7d).
Patch coverage: 12.00% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1623      +/-   ##
==========================================
- Coverage   85.94%   85.50%   -0.45%     
==========================================
  Files         129      130       +1     
  Lines        8098     8155      +57     
==========================================
+ Hits         6960     6973      +13     
- Misses       1138     1182      +44     

Impacted Files	Coverage Δ
src/Distributions.jl	100.00% <ø> (ø)
src/univariates.jl	74.07% <ø> (ø)
src/univariate/discrete/bernoullilogit.jl	12.00% <12.00%> (ø)
src/univariate/continuous/skewnormal.jl	65.62% <0.00%> (ø)
src/eachvariate.jl	90.47% <0.00%> (+4.76%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[devmotion]

Can you mention “parametrized by the log of the odds Logitp = p/(1-p)” in the docstring

Are you interested in the term "log-odds" specifically? The docstring already mentions that it is parameterized by the logit of the success rate, and log-odds is just another name for the logit (see e.g. https://en.wikipedia.org/wiki/Logit).

[devmotion]

Generally this PR should be ready and well-tested.

[matbesancon]

should be good to merge when tests pass I think

[mschauer]

The docstring already mentions that it is parameterized by the logit of the success rate

Well, that requires you knowing what a logit is. Saying it is "parametrised by logitp = log(p/(1-p)) where p is the success rate" makes everything immediately clear to the passerby - doesn't it?

[devmotion]

Ah I see, I thought the main point of your comment was the term log odds. I changed the docstring, hopefully it is clearer now.