Multi-VM benchmark harness: sync client start from a cluster clock and separate op RNG from value RNG

When orchestrating a distributed-systems benchmark over ssh from a laptop (N clients on N VMs, compared across configurations): (1) Compute the synchronized start time from one VM's clock (ssh vm1 'date +%s%3N' + a few seconds) and pass it as a unix-ms flag; the laptop clock is often skewed vs NTP-synced cluster nodes. (2) Use two RNGs per client: one seeded RNG that only decides op type and key, and a separate one for value bytes. Otherwise anything that changes value generation shifts the op/key sequence, and 'same workload seed across variants' silently stops being true. Seed per (run, client) so run r of every configuration replays the same ops. (3) Make the orchestrator resumable by skipping data points whose per-client result files all exist and are non-empty; multi-hour sweeps get interrupted. (4) Stage per-op value files in /dev/shm and time only the RPC so local file I/O doesn't pollute latency.