This is still very much in flow but so far I got, at least locally, most tests running

This refactors the worker state machine such that it follows a similar execution model as the scheduler where we calculate recommendations and messages during a transition and perform these recommended transitions until we converge and there are no further recommendations.

The overall theme of this change is to be less forgiving in edge cases, log more, raise more often. If something unexpected is happening we do not fail silently.
All connected transitions are also linked using a transaction_id which is generated at the top of the chain and propagated through. While I haven't put this into the logs consistently, this is already added to the transition log such that one can easily follow what the reason of a given transition is/was (this is already possible to calculate based on the recommendations but in logs the ID is helpful)

I have currently a strong suspicion that this state machine can be described by only exit and enter actions (hence the few sporadic _transition_enter_{} methods but I haven't converged on this, yet.

One major point about this PR is that I get rid of the release_key and distinguish between delete and forget actions. This helps with keeping state like a suspicious counter and helps with understanding what is actually going on

At the very least the following items are to be finished before this can be considered reviewable

• Get all tests green
• Reiterate signatures of transition functions to accept TaskState instead of key
• Dedicated PR for scheduler changes
• Enable test test_broken_deps (this fails on main as well but is an excellent stress test for the state machine)
• Adress all TODOs / FIXMEs. If not possible, follow up ticket