Add bound to HyperGeometric logp (resolves #4366)

[ricardoV94]

This PR adds a bound to the HyperGeometric logp, in order to return -inf when given invalid parameters, instead of nan or sometimes wrong results (see #4366).

I am using the bounds suggested by @tirthasheshpatel in here, which seem to be the support bounds described in the Wikipedia entry. I am not sure whether the current tt.max and tt.min has any disadvantages compared to the tt.switch that was suggsted before. It seems tt.switch deals better with weird parameter shapes...

To pass the unit tests, it was necessary to replace the nan output of the scipy.hypergeom to -inf. I decided to simply wrap the original scipy function in the helper modified_scipy_hypergeom_logpmf function.

[codecov]

Codecov Report

Merging #4367 (9f9df0e) into master (0ec65e5) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #4367   +/-   ##
=======================================
  Coverage   87.95%   87.95%           
=======================================
  Files          88       88           
  Lines       14491    14493    +2     
=======================================
+ Hits        12746    12748    +2     
  Misses       1745     1745           

Impacted Files	Coverage Δ
pymc3/distributions/discrete.py	95.84% <100.00%> (+0.02%)	⬆️

[tirthasheshpatel]

Everything here looks good to me! Thanks, @ricardoV94!

[AlexAndorra]

LGTM, thanks @ricardoV94 !