Various fixes in fetchAuthEvents

[S7evinK]

This might fix issues with state events gone missing.

[CLAassistant]

All committers have signed the CLA.

[codecov-commenter]

Welcome to Codecov 🎉

Once you merge this PR into your default branch, you're all set! Codecov will compare coverage reports and display results in all future pull requests.

Thanks for integrating Codecov - We've got you covered ☂️

[S7evinK]

Note that I'm happy with adding the failing Sytests to the blocklist, as they test implementation specific behavior.

[kegsay]

No justification for the change, nor any tests.

What's the context here?

[neilalexander]

@kegsay These changes were cherry-picked in from Harmony. Overview:

1. fetchAuthEvents has a causality/ordering issue and is broken today because it doesn't clear the auth set when it fetches a new set of events via federation. The earlier fetched events get incorrectly authed against the few later events that we did know about (if any), which is why so many auth events get unexpectedly rejected with PLs looking like they are trying to reset to an earlier state etc. That bug is now fixed.
2. We need to correctly track the rejected state of auth events through the processing of the event so that fetchAuthEvents can determine whether we need to try to re-fetch those over federation and/or retry them, giving them a "second chance" to be allowed if they were initially rejected for the wrong reason (i.e. due to the above bug).

I didn't write tests and I probably won't at this point because I am happy with the logic of the changes. I've extensively tested them on my own instance and have seen a dramatic reduction in unexpected state resets and/or state going missing, as well as a near-total reduction of unexpected auth errors that aren't signature-related.

This is one of the things we've been missing for literally years now on why Dendrite was unexpectedly unreliable in some rooms with no obvious explanation. It's a huge enough fix to be probably even worthy of a release.