Update IpcStreamFactory state machine to handle being started on a thread that ends

[josalem]

Fixes #39176

"The I/O operation has been aborted because of either a thread exit or an application request."

• ERROR_OPERATION_ABORTED description [emphasis mine]

The first call to IpcStream::DiagnosticsIpc::Listen happens on the thread where EEStartupHelper executes. On Windows, if that thread ends, e.g., in a custom hosting scenario, it causes all Overlapped IO that started on that thread to be cancelled. This specifically includes the asynchronous call to ConnectNamedPipe in IpcStreamFactory state machine. This causes GetNextAvailableStream to go into an error state that would continually check the status of the aborted IO without attempting to reset itself. It always see ERROR_OPERATION_ABORTED and doesn't self-correct. This patch accounts for mid-state change IO cancellations by automatically resetting the Listen port on the error.

I reproduced the behavior from #39176 locally (12% CPU being consumed by the error state) and confirmed that this change fixed the issue (CPU is reported as 0% in the repro).

This patch also updates the write/read error code to only call GetOverlappedResults on infinite timeouts if the error from ReadFile/WriteFile was ERROR_IO_PENDING.

CC @tommcdon @lateralusX

[ghost]

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

[sdmaclea]

LGTM, but I don't completely understand the code.

[sywhang]

LGTM modulo few things 👍

[josalem]

/backport to release/5.0

[github-actions]

Started backporting to release/5.0: https://github.com/dotnet/runtime/actions/runs/346557926