Otherwise, if doc wasn't visited yet, codimd returns some large HTML
page.

It's faster now, makes less joins, doesn't need VALUES.

This brings in an important bug fix so that the output of `-Txdot_json`
is deterministic.

Apparently `dot` has a long lasting bug (recently fixed) which is the
root of the non-determinism:

https://gitlab.com/graphviz/graphviz/-/issues/2510

This was (User, Either CorpusName, [CorpusId]) before, but the case of
UserMaster doesn't make sense with these parameters, so I rewrote the
function to accept only correct datatypes as inputs.

It's not enough to check that the list is empty, we also need to check
that we have some data to begin with, i.e. we need to skip calling
`toFstDate` or `toLstDate` if we have a list of empty strings.

We had CoreNLP hardcoded, however with Spacy this resulted in
selectLems returning no entries, because the algo_id from DB didn't
match. Hence no singulars grouping.

This contains only data, removes some Database dependency in imports.