Generate baseline hashes during tests to reduce flakiness

[ConorMacBride]

• The pytest-mpl test suite relies on lots of hardcoded hashes remaining the same, which leads to lots of test failures on different platforms.
• Instead of hardcoding the hashes we can quickly generate them during each test run and then modify specific hashes to simulate an expected test failure.

Closes #232
Closes #225
Closes #179

[ConorMacBride]

Hi @astrofrog @Cadair! How would you feel about merging this? It would make the test suite much more stable over time, as we'd be less sensitive to minor visual changes in matplotlib. As I mentioned in #240:

I was thinking earlier that we could remove most of the hardcoded hashes in the tests by generating them during the test run: 1) run code to generate hashes 2) modify/break some hashes 3) test that pytest-mpl passes and fails correctly when rerun. It removes the flakiness, but does mean that we only test that matplotlib can produce identical hashes over repeated runs and not necessarily over different matplotlib or OS versions.

[astrofrog]

I'd be happy to see this merged!

[Cadair]

Seems sensible to me.