I.e. do not populate these dict as repo initialization (bootstrap step) but
always use information from database. This is needed because when multiple
instances of the same application run, if one instance adds a CWSource the
other ones will not see it. In particular, when using a scheduler instance,
new CWSource will be added by the web instance and not seen by the scheduler
which is supposed to update them.

We thus define properties for sources_by_eid and sources_by_uri instead
attributes on repository instance. CWSource entities are thus retrieved from
database every time these properties are accessed. We factor out
initialization of the "source" instance (subclass of
cubicweb.server.source.AbstractSource) in a _sources() method. Note that this
method takes care of calling "init" method on the source as well as
"set_schema" (previously done in repo.set_schema(), which now only touches
system_source). Accordingly the "init_sources_from_database" method is dropped
along with "add_source"/"remove_source" methods.

In syncsources hook, we thus drop:

* SourceAddedOp operation which called repo.add_source() so that the
  SourceAddedHook only cares about checking source configuration now;
* SourceRemovedOp and SourceRenamedOp operations for the same reason;
* SourceConfigUpdatedOp as updating the live config of source is
  meaningless once we rely on them being retrieved from the database;
* SourceHostConfigUpdatedHook hook which is now useless without call to
  SourceConfigUpdatedOp;

In 3.10 migration script, remove usage of sources_by_uri repo attribute which,
unless I'm missing something, appears useless (at least now).

In tests:

* unittest_datafeed: remove test_update_url method since we dropped respective
  hook;
* unittest_ldapsource: LDAPFeedUserDeletionTC.test_a_filter_inactivate()
  currently fails because it still relies on live config being updated, this
  will be fixed in the next changeset once all "live source" logic will be
  removed.