In distributed context it is advantageous to use a DataTreeCandidate (resulting from DataTree.prepare()) as a replication unit. Shifting it around to external systems and/or filesystems adds inherent latency, during which we cannot optimistically create another candidate and thus amortize the latency.

This is largely an API problem, as InMemoryDataTreeCandidate has all the state required to prepare a follow-up candidate, which can be committed after it. Expose appropriate APIs so we can take advantage of this.