Fix race condition in System.Threading.Overlapped tests causing IOException on Windows ARM64

[Copilot]

Under PGO stress on Windows ARM64, PackPosTest and PackPosTest1 would intermittently crash with IOException: The handle is invalid at EventWaitHandle.Set() inside the IO completion callback.

Description

Root cause: The two-stage callback chain queues a second NativeOverlapped* that holds the only reference to helper as an unmanaged pointer. The GC cannot trace through unmanaged pointers, so after helper.Wait() returns the JIT is free to report helper as dead — allowing the finalizer to close the ManualResetEvent OS handle before the second callback fires and calls _event.Set().

Fix: Add GC.KeepAlive(helper) after helper.Wait() in both PackPosTest and PackPosTest1:

Assert.True(ThreadPool.UnsafeQueueNativeOverlapped(nativeOverlapped));
Assert.True(helper.Wait());
GC.KeepAlive(helper);  // prevent finalization before second IOCP callback fires

This extends helper's managed lifetime past Wait(), ensuring the ManualResetEvent handle remains valid for the entire callback chain. The bug is latent since the test was written; it surfaced only under defaultpgo stress timing on ARM64.

Fixes #127366

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @VSadov
See info in area-owners.md if you want to be subscribed.