Swap visual tests with hash comparison

[cvanelteren]

Closes #424
As we expand our tests, the runners are running longer and longer. I propose that we move from visual tests to hashes on the runners. The idea would be that we can see what tests fails such that they can be testsed locally to see how they differ visually.

We should move also to nox in case the tests are visually off in a specific combination of python and matplotlib.

[cvanelteren]

Would love your input here @beckermr!

[beckermr]

The slow part here is making the plots. This still makes the plots for both the main branch and the PR.

To get a 2x speedup, we should cache the results from main using the github cache.

[cvanelteren]

Caching is a good idea (see latest commit). I was under the impression that the hash comparison was also faster. The generation on the baseline would be slow, but the hash check is fast.

[cvanelteren]

I will do some speed tests hang on

[beckermr]

The slowest part of this is making the plots. Comparing them via hashes or directly through differencing the images will have a time difference but the change will be small relative to everything else.

[cvanelteren]

Ok yeah great. Another would be to use pytest-xdist which recently got updated to allow to work with pytest-mpl. The latter will give a minor boost as I think there are only 2 cores per runner.