Use a more robust approach to static shape inference in Dirichlet

[brandonwillard]

This PR adjusts the way that static shapes are obtained in the Dirichlet constructor. Previously, in some cases Theano wouldn't properly compute test values for the tt.shape(a) graph. This adjustment simply uses the test value of the a parameter instead.

It doesn't completely finish #4060, because pm.Dirichlet.dist(np.r_[1, 1]).random(size=2) still returns [1., 1.].

[codecov]

Codecov Report

Merging #4061 into master will not change coverage.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #4061   +/-   ##
=======================================
  Coverage   86.83%   86.83%           
=======================================
  Files          88       88           
  Lines       14165    14165           
=======================================
  Hits        12300    12300           
  Misses       1865     1865           

Impacted Files	Coverage Δ
pymc3/distributions/multivariate.py	78.17% <100.00%> (ø)

[AlexAndorra]

LGTM, thanks @brandonwillard !
Out of curiosity, what do you expect pm.Dirichlet.dist(np.r_[1, 1]).random(size=2) to return? Something like [[0.5, 0.5], [0.4, 0.6]] for instance?

[brandonwillard]

LGTM, thanks @brandonwillard !
Out of curiosity, what do you expect pm.Dirichlet.dist(np.r_[1, 1]).random(size=2) to return? Something like [[0.5, 0.5], [0.4, 0.6]] for instance?

Yeah, that would be acceptable.