"Undo" invocation feature

[vigoo]

Currently if a failure happens during the invocation of a worker, it gets automtically retried and if it was not a transient error, eventually the worker gets into failed state.

After that the worker is no longer accessible, and for example if it represents some stateful entity, it is not possible to query it's state (or state before the failing invocation).

We could have a "undo" feature which just drops the last invocation attempt and gets the worker back to the state where it was before the failing invocation.

Of course this is not always safe to do, as there are potentially side effects already performed as part of the failing invocation, but as a user-initiated option I believe it would be helpful.

[vigoo]

Closed because has been split to the following tickets:

• Implement the undo API in worker executor #1180
• Expose the worker undo feature via REST and CLI #1181
• Expose worker undo as a host function #1182