fix(pubsub): always make config check to prevent race (#4606)

In the previous iteration of this PR, #4602, the assumption was made that making the config check earlier will prevent race conditions in the test, which turned out to be incorrect. In addition, although previously ruled out as a solution, this PR makes it so that a config check is always made on the first call to `Receive`. I originally thought this would result in a poor experience for those users who don't have the `roles/pubsub.viewer` permission, but we default `enableOrdering` to `true` anyway if the config call fails due to lack of permissions. Using a default of `true` only negatively impacts those who don't want ordering on their subscription, and this behavior is [already documented](https://github.com/googleapis/google-cloud-go/blob/pubsub/v1.14.0/pubsub/subscription.go#L235-L239).

Fixes #3626

pubsub/subscription.go

@@ -48,7 +48,6 @@ type Subscription struct {
 	mu            sync.Mutex
 	receiveActive bool
 
-	once           sync.Once
 	enableOrdering bool
 }
 
@@ -785,6 +784,8 @@ func (s *Subscription) Receive(ctx context.Context, f func(context.Context, *Mes
 	s.mu.Unlock()
 	defer func() { s.mu.Lock(); s.receiveActive = false; s.mu.Unlock() }()
 
+	s.checkOrdering()
+
 	maxCount := s.ReceiveSettings.MaxOutstandingMessages
 	if maxCount == 0 {
 		maxCount = DefaultReceiveSettings.MaxOutstandingMessages
@@ -897,9 +898,6 @@ func (s *Subscription) Receive(ctx context.Context, f func(context.Context, *Mes
 				}
 				for i, msg := range msgs {
 					msg := msg
-					if msg.OrderingKey != "" {
-						s.once.Do(s.checkOrdering)
-					}
 					// TODO(jba): call acquire closer to when the message is allocated.
 					if err := fc.acquire(ctx, len(msg.Data)); err != nil {
 						// TODO(jba): test that these "orphaned" messages are nacked immediately when ctx is done.