deployment: separate dependencies for backends?

[bertsky]

I'm afraid the heavy toll of packages like ocrd_segment and ocrd_cor_asv_ann (both of which require TF) might not only be annoying here, but quickly lead to clashes between backends.

I understand it would be harder to integrate backends via CLI, but that would make it possible to call them from sub-venvs or Docker, and opt out of them (at least during ocrmultieval setup).

[kba]

Valid points. I am fully open to a different architecture altogether. The background of this project was that we needed to make different evaaluations and I needed an OCR-D processor to provide it. One part that of that was e.g. that we needed to run PRImA LayouEval on Windows and then import those esults in the workspace (not yet in this codebase). So the codebase currently is mostly a proof-of-concept with the goal of "evaluating the evaluation tools".

That being said, if we decide to move foward with such an architecture, we should move every backend to its own package and have a plugin registration mechanism - or even a CLI approach (since we're using filenames anyway) with discovery based on $PATH introspection, so that users can install only those backends they want to use.