[backports-release-1.10] Backport julia/#55910 and julia/#56621

[christiangnrd]

Don't squash?

I backported them while looking into JuliaConcurrent/Atomix.jl#44 and figured I might as well submit a PR.

[IanButterworth]

This is great. Thanks!
Minor thing that helps track backports. Can you make the commit titles include the name of the original PR.

[christiangnrd]

Can you make the commit titles include the name of the original PR.

Done!

[IanButterworth]

Thanks. CI isn't running because this is merging into a backports branch. Should be fixed by #4094

[IanButterworth]

If you rebase this now on the base branch it should run CI

[topolarity]

Thanks for the backport!

For clarity, this will fix the "false cycles" we were adding on accident before JuliaLang/julia#55910, but it leaves in any true cycles that come from the bad "ext → ext" edge definition we have on 1.10

I think that's the right thing to do - just wanted to make it clear what this does / doesn't fix

[topolarity]

I think this might fix a significant majority of our cycle problems on 1.10

Extension circularity should now only occur when two extensions have exactly the same* set of triggers, and the cycles should be a lot shorter / more direct (e.g. "ExtAB ↔ extBA")

* technically the "same" triggers after transitive reduction, meaning not counting any triggers in the dependencies of another