fix(gpu): stop the events gatherer before reinstalling the NVML mock

TestMetricNamesWithoutEBPFProbes calls collectMetricNames twice, and each
call installs a fresh global NVML mock via setupMockDevices. The events
gatherer was stopped from a t.Cleanup, which does not run until the whole
test ends, so the first call's asyncFetchWorker was still reading the
global through EventSetWait while the second call wrote it. The race
detector flagged it in CI (both the go test and bazel jobs run -race);
it did not reproduce locally because I ran without -race.

Stop the gatherer with a defer inside the helper instead. Stop() joins
the worker via wg.Wait(), so the worker is gone before the next setup.

Co-Authored-By: Claude <noreply@anthropic.com>

Author: mbertrone

pkg/collector/corechecks/gpu/nvidia/collector_test.go

@@ -379,9 +379,14 @@ func collectMetricNames(t *testing.T, spCache *SystemProbeCache) map[string]stru
 		testutil.WithArchitecture("blackwell"),
 	)
 
+	// Stop the gatherer before returning instead of registering a t.Cleanup: its
+	// worker goroutine reads the global NVML mock, and the caller invokes this
+	// helper again, which installs a fresh one. A t.Cleanup would not run until
+	// the whole test ends, leaving the first worker racing the second setup.
+	// Stop() joins the worker, so the two can never overlap.
 	eventsGatherer := NewDeviceEventsGatherer()
 	require.NoError(t, eventsGatherer.Start())
-	t.Cleanup(func() { require.NoError(t, eventsGatherer.Stop()) })
+	defer func() { require.NoError(t, eventsGatherer.Stop()) }()
 
 	deps := &CollectorDependencies{
 		DeviceEventsGatherer: eventsGatherer,