Add `runsAfter` relation for activities

[paullatzelsperger]

To express dependencies between activities we currently have the dependsOn relation, which makes Activity B run after Activity A, if B dependsOn A. B does not run if A did not complete successfully, which is acceptable in deployment activities.

However, during compensation activities it might be necessary to order activities, but make them run regardless of whether their predecessor completed successfully.

consequently, we should introduce the runsAfter keyword, so that:

• B runsAfter A -> B runs after A, even if A failed
• B dependsOn A -> B runs after A, only if A succeeded

[jimmarino]

I think this may introduce a lot more complexity and the need to deal with edge cases. Another way to model this is to require agents to always handle failures during undeploy properly. In other words, each agent is always responsible for completing its operations. If an agent encounters an error during undeploy that it cannot resolve, the agent sends a message for manual intervention and proceeds.

This has the same processing effect but keeps states from the orchestration perspective consistent.

[paullatzelsperger]

an easy way to implement this is to always continue the processing if one agent fails during compensation. So consequently, a dependsOn would always mean "run always, even if predecessor fails" on compensation. that way we can make sure that one failing comp does not block subsequent ones.

another alternative way would be to not return an error when a compensation does not succeed, e.g. a participant can't be deleted because it doesn't exist (yet).