DataLoaderTest.EnforcesOrderingAmongThreadsWhenConfigured is flaky

[ezyang]

Sample log:

Nov 14 22:58:37 /var/lib/jenkins/workspace/test/cpp/api/dataloader.cpp:885: Failure
Nov 14 22:58:37 Expected equality of these values:
Nov 14 22:58:37   value
Nov 14 22:58:37     Which is: 1
Nov 14 22:58:37   index++
Nov 14 22:58:37     Which is: 0
Nov 14 22:58:37 [  FAILED  ] DataLoaderTest.EnforcesOrderingAmongThreadsWhenConfigured (1 ms)

Full log: https://circleci.com/gh/pytorch/pytorch/245209?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-build-link

[ssnl]

previously reported at #13634 and #13819 . @goldsborough claimed that it should be fixed in #13919

[goldsborough]

Yes yes I know. I will look

[ezyang]

Another occurrence: https://circleci.com/gh/pytorch/pytorch/247841?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-build-link

[ezyang]

One more: https://circleci.com/gh/pytorch/pytorch/247797?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-build-link

[ezyang]

Looking at CI HUD, it looks like all of these failures occur specifically on the Trusty py3.6 gcc7 configuration; something like a 30% failure rate. It's quite obvious: https://ezyang.github.io/pytorch-ci-hud/build/pytorch-master