Add negative binomial likelihood

[simsurace]

I opted for the p = logistic(f) parameterization. I'm not quite sure whether there are any tradeoffs in choosing this vs. e.g. \mu = exp(f), where \mu is the conditional mean of the negative binomial distribution. The latter parametrization leads to the mean to be only affected by the variance and mean of the latent GP, and not by the likelihood parameter. I haven't thought about whether this has any benefits for parameter learnability that might make it preferable to the p=logistic(f) parameterization.

[codecov]

Codecov Report

Merging #63 (3aa88fb) into master (dc7b997) will increase coverage by 0.39%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master      #63      +/-   ##
==========================================
+ Coverage   94.66%   95.06%   +0.39%     
==========================================
  Files           8        9       +1     
  Lines          75       81       +6     
==========================================
+ Hits           71       77       +6     
  Misses          4        4              

Impacted Files	Coverage Δ
src/likelihoods/gamma.jl	100.00% <ø> (ø)
src/likelihoods/negativebinomial.jl	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update dc7b997...3aa88fb. Read the comment docs.

[simsurace]

Hope the dog is happy now! :)

[simsurace]

Apparently the keyword argument constructor breaks the test for when the second argument logistic is not being wrapped in the link.
How shall we resolve this? Also introduce the keyword constructor for GammaLikelihood and change the tests?

EDIT: Or shall I revert the constructors and we address this in a future PR?

[devmotion]

Change the tests similar to https://github.com/JuliaGaussianProcesses/GPLikelihoods.jl/pull/61/files#diff-d19095b0c530c06b29a86657ed746b5d40a0ebf453a75de8b02aa31b6acb119f, I'd suggest. In particular, it would be good to add tests with the keyword argument and to use also non-default values (e.g., different values for r).

[theogf]

Looks good, I just made some modifications in the docstring, but feel free to reject/modify them.
Can you also bump the patch version?