[MNT] Remove support for Python 3.6, update Windows & macOS VMs, bump manylinux image to manylinux2014

[ulupo]

The time has come to end support for Python 3.6. In fact, this was long overdue. The userbase still on Python 3.6 is almost surely insignificant, and we need to concentrate efforts on shipping for Python 3.10 and ARM architectures, which will considerably increase the CI burden.

In passing, I am updating the Windows and macOS VMs on Azure as follows:

• [MNT] Replace vs2017-win2016 VM with windows-latest as win2016 is deprecated
• [MNT] Migrate to macOS 11 VM as macOS 10.5 is deprecated
• [CI] Replace manylinux2010 with manylinux2014 since SciPy does not support manylinux2010 in 1.8 (see https://github.com/scikit-learn/scikit-learn/pull/23121)
• [Tests] Skip "check_n_features_in" in test_common -- recent versions of scikit-learn require that the n_features_in_ be defined in the fit method, but Binarizer and Inverter do not (and they shouldn't necessarily do so).

[ulupo]

Hey @wreise, would you agree with the proposed changes?

Some tests have started to fail, I think pytest has gotten more adventurous and chooses extremely large parameters -- or some bugs have emerged due to new numpy/scipy versions? Anyway, I think fixing the tests is outside the scope of this PR and would think it's better to quickly open a new PR to fix them.

[wreise]

Hey @ulupo , Thanks for the PR! I'd wait for the CI and see if other typos/something easily-fixable comes up.

[ulupo]

@wreise thanks! Seems we now at least get to the test stage, and they even pass in Windows with 3.8.