Fix ExGaussian logp

[AlexAndorra]

This is a proper version of #4049, with all the changes related to the ExGaussian's logp.
As suggested by @junpenglao in #4045, this PR adds a tt.switch statement in the ExGaussian logp to replace 0 with epsilon. That way, std_cdf never returns 0, and logpow never returns -inf.

The changes seem to work: pm.ExGaussian.dist(0., .25, 1./6).logp(y).eval() doesn't contain -inf anymore, and the model in Discourse doesn't raise a BadInitialEnergy error.

Thanks for the reviews 🖖

[codecov]

Codecov Report

Merging #4050 into master will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #4050   +/-   ##
=======================================
  Coverage   86.81%   86.82%           
=======================================
  Files          88       88           
  Lines       14150    14154    +4     
=======================================
+ Hits        12285    12289    +4     
  Misses       1865     1865           

Impacted Files	Coverage Δ
pymc3/distributions/continuous.py	80.09% <100.00%> (+0.07%)	⬆️

[junpenglao]

Thanks!