De-flake TcpConnectionOwnershipDisposalSpec: keep the close from finishing before the rejected write lands

[Aaronontheweb]

OwnedWrite_ClosingBehaviour_rejection_disposes_owner_before_CommandFailed sends Close, then an owned Write, and expects the Write to be rejected while the connection is closing.

Race: on an idle connection the graceful close completes through self-messages (TransportOperationCompleted, ReadPumpCompleted) within a few ms, and the actor stops. If the test thread is preempted between the two sends, which is realistic on 2-vCPU agents where Close wakes 3–4 threads, the Write is dead-lettered. The handler gets Closed instead of CommandFailed, and the owner is never disposed. Same-sender FIFO doesn't help, because the messages that finish the close come from the actor itself.

• Forced repro: a 20 ms sleep between the sends fails 3/6; 50 ms fails 6/6.
• Natural rate: rare (0/56 full IO-suite runs locally under CPU stress). It was seen once on unmodified dev.

Fix (test only): before Close, park 4 MB of output that the client doesn't read, using 16 KB socket buffers (the same technique as TcpConfirmedCloseSpec). The close can't finish until the test drains it, so the Write always lands in ClosingBehaviour. The test then drains, checks that the blocker arrives intact followed by EOF (the rejected bytes never hit the wire), and expects Closed and Terminated. Every expectation in the test now has an explicit timeout.

Verification:

• 30/30 in isolation.
• Still passes with the forced 50 ms (10/10) and 1 s (3/3) sleeps between the sends.
• Full Akka.Tests.IO filter 5/5.
• Mutation check: removing DisposeOwnedSegments on the ClosingBehaviour rejection path fails it 3/3.
• Also passes on Akka.IO TCP: hold back WriteAck until the output pipe's flush completes #8646's rewritten write path.

The other seven tests in the spec were checked for thread-level races (dispose counts use Interlocked/Volatile, the pool is per test and thread-safe). None were found.