Adding notebook with example of using a "black box" likelihood

[mattpitkin]

This PR adds an example notebook that describes creating a Theano Op for the "black box" likelihood function, including adding a grad method. This has been discussed on the forum here.

The description probably still needs some tidying though!

cc @junpenglao

[junpenglao]

Requested review from @aseyboldt and @ferrine who have more expert theano knowledge ;-)

[ferrine]

Looks very good to me. There is nice motivation, somewhat simple solution (deserves a PR), and a kind warning that this might be a last resort method.

What about the performance of logp_grad function? Could you please additionally profile them in the notebook?

[mattpitkin]

@ferrine At the end of the notebook I've added some tests that check that the gradient function I define is the same as that for the PyMC3 Normal distribution.

Would you like some profiling of the timings of the gradient function?

[junpenglao]

I think a sentence or two pointing to the possibility of doing profiling using http://docs.pymc.io/notebooks/profiling.html is enough.
Also, small formatting nitpick: please follow #2834

[mattpitkin]

@junpenglao I've added profiling of the logpt function, which in my case also seems to profile the gradient, and included a link to the profiling notebook. I've also made some style changes and tried to adopt some of the formatting suggested in #2834. Should I remove the version info for everything other than PyMC3?

[junpenglao]

Looks great like this! Thanks a lot!