Atomically send GOAWAY during HTTP/2 abort

[adityamandaleeka]

Fix a race in Http2Connection.Abort where ProcessRequestsAsync cleanup could call _frameWriter.Complete() between WriteGoAwayAsync and Abort(), causing GOAWAY to be dropped.

Add Http2FrameWriter.AbortWithGoAway() and use it from Http2Connection.Abort() to perform GOAWAY write/flush and abort under one lock acquisition.

I'm intentionally not adding a regression test for this because the amount of timing nonsense you have to do to simulate it defeats the purpose of the test. I noticed this happening in CI though so hopefully we'll see less flakiness here.

[Copilot]

Pull request overview

This PR fixes a race condition in HTTP/2 connection abort handling where a GOAWAY frame could be lost. The race occurred when ProcessRequestsAsync cleanup called _frameWriter.Complete() between the WriteGoAwayAsync and Abort() calls, causing the _completed flag to be set to true and the GOAWAY frame to be discarded.

Changes:

• Introduced Http2FrameWriter.AbortWithGoAway() method to atomically write GOAWAY frame and abort connection under a single lock acquisition
• Modified Http2Connection.Abort() to use the new atomic method when closing the connection for the first time
• Added detailed documentation explaining the race condition and the fix

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
src/Servers/Kestrel/Core/src/Internal/Http2/Http2FrameWriter.cs	Added new AbortWithGoAway() method that atomically writes GOAWAY frame and performs connection abort under single lock to prevent race condition
src/Servers/Kestrel/Core/src/Internal/Http2/Http2Connection.cs	Updated Abort() method to use new atomic AbortWithGoAway() when TryClose() succeeds, and regular Abort() otherwise

[adityamandaleeka]

ha, this ran into the flakniess fixed in #65456

[dotnet-policy-service]

Looks like this PR hasn't been active for some time and the codebase could have been changed in the meantime.
To make sure no conflicting changes have occurred, please rerun validation before merging. You can do this by leaving an /azp run comment here (requires commit rights), or by simply closing and reopening.