Implement merging dataset and preemptive reading at the same time

[gullmar]

Example: you want to start multiple runs and read a bunch of their output items as soon as they are available, e.g., 100 items from the 2nd run, 100 items from the 5th one, and so on.
Now, what if the actor is migrated? We should guarantee that the order in which items are read after the resurrection is the same, if runs are persisted.

My idea would be to persist the order in which datasets are accessed for a specific group of runs.

The way we achieve this may depend on issue #4.