macOS-only failures in CI with RDKit visualization

[mattwthompson]

Log: https://github.com/openforcefield/openff-toolkit/runs/5504391595?check_suite_focus=true

Traceback:

________________ TestMoleculeVisualization.test_visualize_rdkit ________________
openff/toolkit/tests/test_molecule.py:3797: in test_visualize_rdkit
    assert isinstance(mol.visualize(backend="rdkit"), IPython.core.display.SVG)
openff/toolkit/topology/molecule.py:6479: in visualize
    from rdkit.Chem.Draw import (  # type: ignore[import]
../../../miniconda3/envs/test/lib/python3.9/site-packages/rdkit/Chem/Draw/__init__.py:14: in <module>
    from rdkit.Chem.Draw import rdMolDraw2D
E   ImportError: dlopen(/Users/runner/miniconda3/envs/test/lib/python3.9/site-packages/rdkit/Chem/Draw/rdMolDraw2D.so, 2): Library not loaded: @rpath/libfontconfig.1.dylib
E     Referenced from: /Users/runner/miniconda3/envs/test/lib/libcairo.2.dylib
E     Reason: Incompatible library version: libcairo.2.dylib requires version 14.0.0 or later, but libfontconfig.1.dylib provides version 13.0.0

I can't reproduce this locally (macOS, Python 3.9) with the same versions of cairo, pycairo, and rdkit, the latter of which was recently bumped to 2021.09.5. I checked the RDKit and its feedstock's issue tracker and nobody else has reported it yet.

[mattwthompson]

A similar issue was reported and fixed by the maintainer of rdkit-pypi: kuelumbus/rdkit-pypi#33

[mattwthompson]

I somehow did not have the thought of looking as far as the fontconfig feedstock, where I found a familiar face already tracked this down and reported it! conda-forge/fontconfig-feedstock#54

[mattwthompson]

Looks like everything's been fixed upstream. No changes to our code needed. I kicked the tires on it in #1225 and the import no longer causes a crash. If this causes deployment issues in the future we can deal with them in packaging.

[mattwthompson]

I'm seeing this pop up again. It seems like the problem is when an old build of cairo is pulled down, like cairo-1.16.0-he69dfd1_1008. It looks like the version (1.16) has stayed the same for some time but is periodically rebuit, i.e. the new builds are cairo-1.16.0-something_1010. I can't figure out a way to do i.e. >=1010; as far as I can tell the build string is processed like a string, not a number/version, and may have other weird things going on.

If this continues to be an issue, perhaps other upstream changes are needed to ensure RDKit isn't installed alongside one of these builds.

[mattwthompson]

Seems to actually be fixed after conda-forge/conda-forge-repodata-patches-feedstock#248