DistributedData multi-node: de-flake ReplicatorSpec (per-attempt probes for Get reads)

[Aaronontheweb]

ReplicatorSpec reads CRDTs with a 50 ms ExpectMsg inside AwaitAssert on the shared TestActor, so one reply later than 50 ms desyncs the queue (CI: GetSuccess(D1:GCounter(1)) did not match, or a stray GetSuccess(KeyC) that breaks the next step). Each attempt now uses a fresh probe, the same fix #8509 made for ReplicatorChaosSpec.

Changes

• ReplicatorSpecTests and every Cluster_CRDT_should_* step are now async Task, awaited in order. They use RunOnAsync, ExpectMsgAsync, ExpectNoMsgAsync, AwaitAssertAsync, WithinAsync and EnterBarrierAsync (new JoinAsync and EnterBarrierAfterTestStepAsync helpers), so no step blocks a pool thread with sync-over-async. Assertions and values are unchanged.
• The Get-read loops (KeyC, D1..D30, KeyA, replica counts) use AwaitAssertAsync with a fresh TestProbe per attempt, a 1 s AttemptTimeout and a 200 ms interval. The converge step sends all 30 Gets, then reads the 30 replies in order. Overall budgets (5 s, 10 s) are unchanged.
• TestConductor.Blackhole/PassThrough are awaited instead of .Wait(...), so a failed partition change fails the test.
• The three KeyB loops in the 2-node step get a 10 s budget (was the 3 s default, equal to the inner expect), with a fresh probe per attempt.

The Cluster_CRDT_should_support_prefer_oldest_members step is converted too, but ReplicatorSpecTests does not call it.

Testing

Builds clean with -warnaserror. On head 1aa39e9, all four multi-node jobs (Linux, Linux Artery, Windows, Windows Artery) and the unit-test jobs pass in CI.

Breaking changes

None (tests only).