Special-case handling of maintenance-style RPCs
Right now we're explicitly replicating maintenance-style RPCs to all nodes. This introduces a lot of complexity into the replication queue given that 9 of 12 replication types are for maintenance jobs. And ultimately, this is not work that is really worth it given that a repository state doesn't depend on whether a repository is well-packed or not.
We should rework the way we execute maintenance RPCs to instead use a best-effort strategy: instead of replicating, we just execute it on all nodes which are currently online. Note that "online" doesn't even need to take into account the current repository generation given that it doesn't matter whether a node is behind, it can execute replication jobs regardless of that.
Edited by Patrick Steinhardt