only use multiprocessing if max_workers > 1 (partial revert of f95111ff)

[bertsky]

This was a self-own: in f95111f I tried to be smarter than my past self and changed the criterion when to use the ProcessPool executor –

• from max_workers > 1 (i.e. whether OCRD_MAX_PARALLEL_PAGES was requested and the processor implementation supports that)
• to isinstance(workspace.mets, ClientSideOcrdMets) (i.e. whether the workspace can be processed in parallel)

But for such important cases like Tensorflow, where (unless you put the model in a singleton background process connected via queues to the page workers) multiprocessing is impossible (because the CUDA context cannot be shared), this is clearly wrong. We have to be able to prohibit in the processor implementation (via max_workers = 1) multiprocessing.