Race condition in plugin registration

[bertinatto]

What happened:

A race condition in the plugin watcher might happen when a pluginwatcher.Watcher is created, started and immediately stopped and cleaned up. If executed several times, the unit test TestNewManagerImplStartProbeMode will expose this issue.

The PR #71440 addressed part of the problem by serializing the handling of CREATE and DELETE events. However, a race might still happen in the pluginwatcher.traversePluginDir function. #71440 (comment) describes in detail the problem.

How to reproduce it (as minimally and precisely as possible):

# Create a binary
$ go test -c k8s.io/kubernetes/pkg/kubelet/cm/devicemanager -run  ^TestNewManagerImplStartProbeMode$

# Run the test in a loop until it fails
$ while ./devicemanager.test -test.failfast=true ; do :; done

/sig storage
/sig node

[bertinatto]

/assign