WIP: Add Python 3.8 builds

[grlee77]

I recently made a release for PyWavelets supporting Python 3.8. I thought I would try applying similar changes here while I still recall what was needed.

Mainly, this PR updates multibuild to a newer commit that can install Python 3.8 for the user on Appveyor if needed.

[hgomersall]

I appreciate the effort! It doesn't seem entirely successful on CI yet though...?

[grlee77]

Yeah, I didn't get a chance to come back to it yet. I did get it working for PyWavelets, so there is probably just a subtle difference here that is causing the failure.

[hgomersall]

Regardless, I appreciate the effort!

[grlee77]

I think multibuild was still on the wrong commit last time. That should fix the Python 3.8 setup issues.

However, I think the other difference here vs. PyWavelets is that these builds use SciPy while the PyWavelets ones did not. I don't think SciPy has wheels for Python 3.8 yet, so it will probably be easier if we wait for that.

Also, I noticed in some cases pip installed SciPy 1.3.1 with an older NumPy like 1.10.4 or 1.12.1 which is incompatible. To resolve that we can either bump NumPy to 1.13.3 (from Sept. 2017) in these builds or pin SciPy to appropriate versions instead of allowing pip to pull the latest one.

[grlee77]

Actually only the Windows builds try to install SciPy due to the inclusion of:
- pip install -r requirements.txt in .appveyor.yml

The Travis script never uses the requirements.txt file, so SciPy gets left out. I guess we were not currently running the test suite on linux/OS X, but should probably add that at some point.

[hgomersall]

I'm confused - so you think the scipy parts of the test suite are being ignored on Travis?

[grlee77]

I'm confused - so you think the scipy parts of the test suite are being ignored on Travis?

As far as I can tell no tests are being run at all on Travis (They are at the main repo, but not here on these wheel builds). I am not sure why that was missing.

I see there is #11, but that seems to be adding support for running the FFTW library tests, not the pyFFTW ones.

[grlee77]

It looks like the windows case for 3.8 should start working once SciPy has wheels. Currently it fails to build from source because we do not try to install BLAS/LAPACK here. I would recommend to just wait on SciPy before proceeding further.

[grlee77]

Trying this again now that SciPy has wheels for 3.8

[grlee77]

Okay, this is looking good. The only test failures are the Dask-related ones reported in pyFFTW/pyFFTW#276

[grlee77]

these commits are now in #30 for the new release, so I will close this one