CI: Speeding up conda solver time by removing pymapd and pinning some packages

[datapythonista]

xref #2181

Running conda env create --name ibis-dev --file ci/requirements-3.7-dev.yml is currently taking 36 minutes (for the conda solver). This causes the builds to reach the 1 hour limit and fail (the docs, benchmarks and conda build builds).

Pinning pymapd reduces the time to to 20 minutes (from the 36). We still need a better solution (20 minutes for the solver is crazy), but at least this should fix the CI in an easy way for now.

[jreback]

let’s remove pymapd entirely for now

this broke a lot of things

[datapythonista]

Removed the dependency here, so we can see the change in the solver speed in the CI for now. I guess we need to disable the backend from the tests too. Anything else?

[datapythonista]

@jreback this solves the main issue with the conda solver being too slow. The remaining errors are the ones caused by not having enough space in disk. And seems like the LinuxBuildConda has another set of conda dependencies that are also too slow. I'll have a look to that in a separate PR.

If we can get this merged, that would be very helpful. So, I can finish the disk space problems without conda being too slow. Thanks!

[jreback]

thanks @datapythonista IIRC @xmnlab has some patches for the image size. though maybe we just want to split the testing in 2

[datapythonista]

I'm testing on splitting the builds in two in #2194, and seems quite simple. Will re-run now that PR with this fixed, and see how it looks. I think just two are missing if we go for the splitting approach:

• See what to do with doctests, which some will fail if the backend is not set up (can't skip doctests with the marker as the rest)
• The long conda solve in the LinuxBuildConda build