`test_meg_prob` seems to test multiple things at once and is hard to follow

[pmeier]

Maybe split it into multiple smaller ones? Rule of thumb should be to only have a single assert per test.

[a-ws-m]

I've split up the test somewhat, added comments that better explain the components, and added some static, partially trained models that facilitate the unit tests. There are a couple of examples of integration tests, whereby the setup for one test directly enables a follow-up (for example, checkpoint files should be generated during training, and saving should work with checkpoint files, so saving after a training routine is a reasonable way to test this).