Update ppc test

[ColCarroll]

I think this will unstick @StephenHogg in #4211.

My working theory is that this was asserting that 2 arrays of 20,000 floats were equal, after going through a couple arithmetic operations, and so it was essentially fuzzing theano's math library. I can't prove that, but this

• makes the fuzzing less thorough (2,000 floats now), and
• lowers the number of operations (I don't go through a pandas dataframe in the middle, since I feel like that might do wild things to the floats).
• Added an extra seed before each ppc sampling, since two of the asserts should give the exact same results.

Also, I changed assert np.allclose to npt.assert_allclose, since it gives more pleasant stack traces.

[codecov]

Codecov Report

Merging #4246 (bd7e1d9) into master (b990e49) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #4246   +/-   ##
=======================================
  Coverage   88.19%   88.19%           
=======================================
  Files          89       89           
  Lines       14358    14358           
=======================================
  Hits        12663    12663           
  Misses       1695     1695           

[Spaak]

LGTM; although I'm not sure why the pandas dataframe was there to begin with?

[ColCarroll]

Also not sure. Maybe for coverage? It is still in a few other tests.