merge with stable

3.15 is the new old stable

[web] don't link to None in author box (when author has been deleted) (closes #2409855)

dbapi: try to restore compatibility of the connect api with 3.15 A few issues here: - urlparse('foo').scheme is '', not None - cnxprops is an optional argument Regression from 62213a34726e ("[db-api/configuration] simplify db-api and configuration so that all the connection information is in the repository url, closes #2521848") Closes #2754322

3.16 is the new stable After discussion with David Douard I'm merging 3.16.x branches in stable and starting 3.17 feature on default.

Added tag cubicweb-debian-version-3.15.10-1 for changeset feca12e4a618

Added tag cubicweb-version-3.15.10 for changeset 89bdb5444cd2

prepare 3.15.10

Added tag cubicweb-debian-version-3.16.1-1 for changeset 84fbcdc8021c

Added tag cubicweb-version-3.16.1 for changeset d95cbb7349f0

prepare 3.16.1

[views/primary] some inner sections should use the `limit` by default to avoid a denial of service (closes #2719110) Today, it is possible to call .related and get a huge unlimited database-dos-inducing resultset that will be nevertheless limited a bit further in pure python in the `autolimited` view. While we cannot completely avoid potential denial of services such as these we mitigate the problem with the default ui settings: if the inner vid is `autolimited`, then the relation result sets is computed using the user-defined limit. This change respects the semantics of the `autolimited` view and shouldn't break anything.

[entity] ensure the .related(entities=False) parameter is honored all the way down (closes #2755994) As of today, such a call will always fill the relation cache by calling .entities() on every single related rset entry. As a consequence, the `limit` parameter handling also had to be fixed. It was bogus in the following ways: * not used in the related_rql, hence potentially huge database requests, but also actually * foolishly used in the .entities()-calling cache routine we now bypass (this changeset ticket's main topic) Now: * we set a limit on the rql expression, and * forbid caching if given a non-None limit (as we don't want to make the cache handling code more complicated than it is already) With this, entity.unrelated gets a better limit implementation (so the code in related/unrelated is nice and symmetric) Risk: * _cw_relation_cache disappears completely, which is good, but this is Python, so you never know ...

[test/web] fix invisibly bogus test (prepares #2755994) The test was wrong but that was cancelled out by a cache effect and fuzzy naming. Wiping the entity caches restores sanity: the choices list are the same before and after the SET. Also field.choices uses entity.unrelated but always returns related + unrelated elements. Hence `choice` replaces `unrelated` where it makes sense. AssertIn is used in place of AssertTrue.

[cwconfig] Fix exception handling when building the cube dependency graph UnorderableGraph exceptions do not have a 'cycles' property. A simple cast to str does the job.

[merge] backport stable fixes