The logic in the initial commit (97dd107070) was broken in some cases,
and would end up tagging as unused IDs that had valid usages. It is
reverted by this commit..
For this new solution to #98029 (deleting unused archipelagos of data),
the logic handling dependency loops detection is reworked to rely on a
new tag in the relations entry of the relevant IDs, instead of
pre-tagging as unused the ID itself.
Further more, when a dependency loop is detected, the IDs in it cannot
be immediately tagged as unused, since it may be that the entry point
of that loop is later detected as actually used. So their relations
entries are not tagged as processed, and only the entry point of the
potential archipelago can be checked in that case, outside of the
recursive processing of dependencies.
The other IDs of the archipelago will then be processed again later, in
a state where at least one ID in the archipelago has a known state for
sure, which then allows for a safe evaluation of the other related data.
This commit should be backported to 3.3LTS.
ac1ac6be9a