De-flake ClusterShardingLeaseSpec: join the cluster in InitializeAsync instead of blocking a pool worker in the constructor

[Aaronontheweb]

What changes

ClusterShardingLeaseSpec, the base that all fifteen lease tests and variants run through, no longer joins the cluster from its constructor with a blocking wait. The join moves into InitializeAsync and awaits Cluster.JoinAsync under a dilated 30 s cancellation bound, which completes on the node's own MemberUp and parks no thread. Sharding starts with StartAsync. A DisposeAsync bridge calls Dispose() so xUnit v3 still runs the TestKit teardown; once #8545 lands this becomes an override that chains to the base, and a comment says so.

The facts pass TestActor explicitly on every region tell, because the implicit sender reads a thread-static cell and a fact resumes on a different thread after InitializeAsync. The nested expect inside the lease-lost recovery loop gets a 1 s inner budget so the outer 10 s loop retries about ten times instead of three. The whole file moves to the async TestKit API per the maintainer's rule: every fact is async Task, and every wait is the Async form. No assertion changes.

Why

Build 131196 (Windows unit tests) failed PersistenceClusterShardingLeaseSpec.Cluster_sharding_with_lease_should_recover_if_lease_lost, but not in the lease logic. The constructor asserted that the node had reached Up inside a blocking AwaitAssert with a flat 3 s default, polling every 100 ms. MemberUp arrived at 3.462 s, a 462 ms miss, after 3.17 s of log silence. The cluster daemon's startup needs about six dispatches on the default dispatcher, which is the thread pool, and the constructor's blocking wait held one of the two workers a 2-vCPU agent starts with. xUnit v3 never raises the pool floor for this assembly because it disables parallelization. The lease, shard-stop, and persistence paths have no race; the analysis ruled each out.

A product item from the analysis is not in this PR: Shard.ReleaseLeaseIfNeeded blocks in PostStop with a synchronous wait where Pekko fires and forgets.

How it was checked

dotnet build src/contrib/cluster/Akka.Cluster.Sharding.Tests -c Release -warnaserror clean. All fifteen tests and variants run three times: 15 passed each time. A grep for the synchronous TestKit forms over the file returns nothing. BREAKING_CHANGES_V1.6.md untouched; this is test-only.

Second commit

From the adversarial review: the _ = TestActor touch in InitializeAsync was redundant, because the ambient-context attribute already pins the actor cell on the test thread before each fact, and the touch only left a live cell on a returning pool worker. Removed with its comment. The spec passes 15 of 15. #8545 is stacked on this PR and converts this spec's lifecycle methods to overrides once the TestKit gains the virtuals.