De-flake ORMultiValueDictionary_WithValueDeltas_LargeDataSet: replicate with WriteLocal and poll the replicas

[Aaronontheweb]

ReplicatorSpecs.ORMultiValueDictionary_WithValueDeltas_LargeDataSet failed on the AOT stack's CI (#8604, Windows unit tests) with UpdateTimeout on its second WriteMajority update. The failure has nothing to do with that PR. This is the third de-flake of this test. The first two (#5686, #7899) raised the timeout; this one removes the cause.

Cause

• ORMultiValueDictionary.SetItems clears and re-adds every element (the same as JVM ORMultiMap.put). Each write in this test ships the full 20,000-element set, delta included.
• With WriteMajority, ReadWriteAggregator resends the full state to the primary and secondary after Timeout/5, which is 2 s here. On a 2-vCPU agent one hop already took about 2.5 s. So each write sent roughly three full payloads between three in-process ActorSystems sharing one thread pool.
• The failing run's log shows heartbeats 2–5 s late, nodes flapping UNREACHABLE, and WriteAcks arriving 5–9 s after the aggregator had timed out.
• Reachability does not gate WriteMajority; the flapping was a symptom of the CPU load.

Change

• Both updates now use WriteLocal.
• The test polls each replica's local copy (Get with ReadLocal, inside AwaitAssertAsync) until the entries equal the expected set, with a Dilated(30s) bound. The data still replicates across all three nodes, through delta propagation (every 400 ms) and gossip.
• The synchronous ExpectMsg<Changed> calls, each of which blocked a pool thread for up to 10 s, are gone.

The test still catches the #5663 quadratic SetItems/merge regression it was written for. It no longer covers WriteMajority with a 20,000-element payload; the rest of ReplicatorSpecs still covers WriteMajority.

Verification (local)

• 20/20 passes at 7–11 s each.
• 10/10 passes at 10–15 s each, with every core pegged by busy loops.
• The full ReplicatorSpecs class passes, 7/7.