TS-4507: Fix SSN and TXN hook ordering.

[shinrich]

This change was motivated by dealing with ordering problems between SSN_CLOSE and TXN_CLOSE causing problems (crashes). This we addressed as follows

• The TXN_CLOSE is kicked off in the State Machine kill_this. This is the point were we know that the TXN really is going away. This is unchanged. What this patch adds is a call to transaction_done to let the ClientTransaction know that TXN_CLOSE has completed so the calculation about whether is it time to execute SSN_CLOSE or not. As such we cannot null out ua_session early.
• To support SSN_CLOSE accurately, we split the destroy logic into two parts: destroy and free. Destroy commits to deleting the session and it kicks off the SSN_CLOSE plugin. free performs the final resource reclamation of the session object.

In addition this PR includes the following fixes.

• In ProxyClientSession::state_api_callout we schedule_in 10ms in the future if the plugin lock is not acquired. Saw ASAN use-after-free crashes when the Http2ClientSession is deleted but the schedule event remains and is triggered. Added a schedule_event member to track this case and cancel any outstanding schedule events on free.
• Cleaned up handling regular events at the same time as plugin events. The original code relied on the subclasses overriding handle_api_event to handle the regular events, but the handler only handled the TIMEOUT event. Changed that to augment the subclasses' main event handler to call out to state_api_callout in the event of the plugin events.

[atsci]

FreeBSD build successful! See https://ci.trafficserver.apache.org/job/Github-FreeBSD/407/ for details.

[atsci]

Linux build successful! See https://ci.trafficserver.apache.org/job/Github-Linux/301/ for details.

[shinrich]

These fixes improved the stability of our 5.3.x + TS-3612 build. I'm heading out until July 14, but I wanted to make these changes available in case anyone else was seeing stability problems in this area.

[jpeach]

Is is possible to break the addditional fixes out into separate tickets and PRs?

[shinrich]

It would be possible to break out the schedule_event clean up and the state_api_callout into smaller patches. The rest of it really needs to be together. I need to move onto another project by Monday. I'll see what I can get done this afternoon. Otherwise these fixes will need to wait a while.

[shinrich]

Pushed new version of the branch that does not include the schedule_event and state_api_callout fixes. Filed TS-4663 and TS-4664 to track those issues. Will try to reproduce those fixes on new branches.

[atsci]

FreeBSD build failed! See https://ci.trafficserver.apache.org/job/Github-FreeBSD/432/ for details.

[atsci]

Linux build successful! See https://ci.trafficserver.apache.org/job/Github-Linux/325/ for details.

[SolidWallOfCode]

I've reviewed the changes and I don't think they can really be split up. Part of the reason is similar changes in both the HTTP/1 and HTTP/2 classes, which makes the change appear bigger than it is. I'm +1 on committing this, especially since we've been successfully testing them in production inside Yahoo!.