AsyncLazy bugfix

[majocha]

Description

New code, new bugs:
#18074 (comment)

It is incredibly easy to write buggy code when concurrent mutations of state are involved. 😐
This change is to contain the mutation to just a single function with lock.
Fixes this rare edge case with unawaited completion in AsyncLazy:
#18074 (comment)

[github-actions]

⚠️ Release notes required, but author opted out

Warning

Author opted out of release notes, check is disabled for this pull request.
cc @dotnet/fsharp-team-msft

[majocha]

@0101, I couldn't repro locally this timeout in stress-test you pointed out:
https://dev.azure.com/dnceng-public/public/_build/results?buildId=883746&view=ms.vss-test-web.build-test-results-tab&runId=23109844&resultId=100187&paneView=debug
but I fixed another bug, so maybe possibly it will help with it?

Maybe it would be better to use --blame-hang-timeout in the CI instead, with some generous timeout, like 5 minutes? In case of a genuine deadlock or hang we'd at least get a dump.

[0101]

@0101, I couldn't repro locally this timeout in stress-test you pointed out: https://dev.azure.com/dnceng-public/public/_build/results?buildId=883746&view=ms.vss-test-web.build-test-results-tab&runId=23109844&resultId=100187&paneView=debug but I fixed another bug, so maybe possibly it will help with it?

Let's try!

Maybe it would be better to use --blame-hang-timeout in the CI instead, with some generous timeout, like 5 minutes? In case of a genuine deadlock or hang we'd at least get a dump.

Yeah that sounds like a good idea. Hopefully we don't have any actually long running test. Well if we do we'll find out.

[psfinaki]

Thanks! We are getting there... :)