Fix DCGM group cleanup on watch and health error paths

[rvatkar]

Summary

Fixes DCGM group lifecycle issues:

• DCGM-6901: destroy health-check groups on error from healthCheckByGpuId.
• DCGM-6902: destroy PID-watch groups on error from watchPidFields.
• DCGM-6903: destroy field-watch groups on error from WatchFields.
• DCGM-6917: return UpdateAllFields errors from WatchFields and watchPidFields, matching WatchFieldsWithGroupEx.
• DCGM-6921: destroy the REST sample PID-watch group after getProcessInfo.

The watch helpers now clean up only on failure and preserve group ownership on success. PID-watch DCGM errors still return *dcgm.Error, and the conditional getSupportedDevices() path is unchanged.

Tests

• go test ./pkg/dcgm -run 'Test(GroupCapErrorMatcher|HealthCheckByGpuIdDoesNotLeakGroupOnError|WatchFieldsDoesNotLeakGroupOnError|WatchPidFieldsDoesNotLeakGroupOnError| WatchPidFieldsDoesNotLeakGroupOnNonRootWatchError|WatchFieldsDoesNotDestroyGroupOnSuccess|WatchPidFieldsDoesNotDestroyGroupOnSuccess|WatchFieldsReturnsUpdateAllFieldsErrorAndDestroysGroup| WatchPidFieldsReturnsUpdateAllFieldsErrorAndDestroysGroup)' -count=1 -v
• go test ./samples/restApi/... -count=1

[nccurry]

Applies to tests/processinfo_test.go:21-23, which is outside this PR diff:

go-dcgm/tests/processinfo_test.go

Lines 21 to 23 in a47f5c5

	group, err := dcgm.WatchPidFields()
	if err != nil {
	t.Fatalf("Failed to watch PID fields: %v", err)

The PR added non-root-aware coverage for the new package tests, but the existing process-info tests still hard-fail when WatchPidFields returns Host engine is running as non-root. Because this PR changes PID-watch cleanup, go test ./... cannot validate the repo on the same non-root environment where the new regression tests now skip correctly. Consider applying the same root-aware skip/helper here.