[0x000NULL]

Fixes #4322

On Windows, the subtests of TestNetworkInspectWithContainers each create nat networks and start a container on them, all in parallel (tigron runs tests in parallel unless NoParallel is set), together with the rest of the network package tests. One of them regularly fails with:

plugin type="nat" failed (add): hcnCreateNetwork failed in Win32: The object already exists. (0x1392)

#4360 added a one second sleep after network create. That did not fix it: the baseline runs below were made with the #4360 sleeps in place and still failed.

This sets NoParallel on the test and its subtests when running on Windows (NoParallel is not inherited by subtests, so each one needs it; this follows container_run_network_windows_test.go), and removes the sleeps in this test. Other platforms are unchanged. I left the sleep in TestNetworkInspectDualStack alone; it is skipped on Windows.

I have not identified the exact HNS race, only that serializing these tests makes it go away.

Testing

Runs on a fork (windows-2022, same job-test-in-host workflow, trimmed to the windows job):

Before (main):

• https://github.com/0x000NULL/nerdctl/actions/runs/38935270736: full suite, 10 jobs, 2 failed with 0x1392 in TestNetworkInspectWithContainers
• https://github.com/0x000NULL/nerdctl/actions/runs/38938321572: only TestNetworkInspectWithContainers, 40 jobs, 2 failed with 0x1392
• main plus only the separate cni.sh fix, full suite, 5 runs x 10 jobs: 12 failed with 0x1392 in TestNetworkInspectWithContainers (runs 38944446472, 38944456629, 38944460822, 38944464805, 38944468969)

With this change:

• https://github.com/0x000NULL/nerdctl/actions/runs/38940150149: full suite, 10/10 passed
• https://github.com/0x000NULL/nerdctl/actions/runs/38940142053: only TestNetworkInspectWithContainers, 60/60 passed
• full suite, 4 runs x 10 jobs (runs 38952948389, 38952961281, 38952965325, 38952969222): no 0x1392; 38/40 passed, the 2 failures were an unrelated TestRunRmTime timing flake

So 16/100 jobs hit 0x1392 without this change and 0/110 with it. The runs with this change also include the separate cni.sh fix (#5268), which does not touch this test.

go vet passes for the package on windows and linux, and gofmt is clean.

[AkihiroSuda]

Thanks