The benchmarking tests were failing because of the incorrect
implementation of the ScanAllByPrimaryKeyCursor. The previous
implementation caused the currently allocateable 1m edgeids to run out
very quickly, causing the the tests to freeze.
This PR adds support for generating randomized workloads that will be executed
against a simulated cluster, as well as against a correctness model. Initially
this just generates ScanAll and CreateVertex requests, and anything that it
creates, it also inserts into a `std::set`, and when we do a ScanAll, it asserts
that we get the same number of requests back. This will become much more
sophisticated over time, but it's already hitting pay-dirt.
