Build test suite to detect changes that break loading of models

[vpratz]

To avoid accidentally breaking serialization/model loading, we need an automated testing framework. It should store checkpoints and outputs produced with one version, and load the checkpoints and compare the outputs in a different version.

For maximum flexibility, it would probably best if we could run it on arbitrary git refs, to allow tracing down potential breakages and checking it locally during development.

This will be a larger undertaking, but should have priority for version 2.1 so that we can make serialization more stable.

[LarsKue]

Assigning myself because I will most likely tackle this, but if anyone else wants to give it a go, let me know!