Retain subject IDs when loading vocabulary over existing one

[osma]

Currently if you load a new version of a vocabulary over an existing one, what likely happens is that the integer subject IDs used internally within Annif will change. This will cause mismatches with backends that use the subject IDs (e.g. tfidf, fasttext, vw_multi) within their models, so you have to train them again.

Instead we could try to match URIs in the new vocabulary with old subject IDs and reuse them as much as possible. Then at least loading a vocabulary with additional concepts would not break existing models. If concepts have disappeared from the new version of the vocabulary, they should be marked as nonexistent and filtered from suggestion results.