Unpin theano

[ColCarroll]

Another try to unpin theano, since it appears nothing is building, maybe enough restarts will help this one.

[junpenglao]

E       raise RuntimeError('To use MKL 2018 with Theano you MUST set "MKL_THREADING_LAYER=GNU" in your environement.')

[twiecki]

We probably shouldn't be blocked by travis. If tests pass locally I think we should merge.

[junpenglao]

(From @ColCarroll)
i cannot recreate the failures locally — if anyone is running linux and can, that would be interesting!
python -m pytest -vsx pymc3/tests/test_sgfs.py::test_minibatch
python -m pytest -vsx pymc3/tests/test_text_backend.py::TestTextSampling::test_supports_sampler_stats

I found out what the problem is: both test fail if setting njobs=1, which is the default for some reason in Travis.
Still trying to find out why test fail when njobs=1 tho, my guess is something to do with the random_seed

[ColCarroll]

Wow, thanks for fixing!

I think this is good enough for now – one is still marked experimental, one is not super well maintained. Seems like two reasonable goals for the next release might be making sgfs non-experimental, and deprecating all backends except h5py.

Will let someone else merge, or object!

[nouiz]

We think we fixed the problem, but as we where not able to reproduce it, can you test with the master of Theano to confirm if the batch_gemm compilation problem is fixed?