grpclb: consider IDLE SubConns as connecting (grpc#4031)

Otherwise, when the first response is received from the grpclb server, the
parent ClientConn enters TransientFailure, and the first several
non-wait-for-ready RPCs will fail.

balancer/grpclb/grpclb.go

@@ -288,7 +288,11 @@ func (lb *lbBalancer) regeneratePicker(resetDrop bool) {
 //
 // The aggregated state is:
 //  - If at least one SubConn in Ready, the aggregated state is Ready;
-//  - Else if at least one SubConn in Connecting, the aggregated state is Connecting;
+//  - Else if at least one SubConn in Connecting or IDLE, the aggregated state is Connecting;
+//    - It's OK to consider IDLE as Connecting. SubConns never stay in IDLE,
+//    they start to connect immediately. But there's a race between the overall
+//    state is reported, and when the new SubConn state arrives. And SubConns
+//    never go back to IDLE.
 //  - Else the aggregated state is TransientFailure.
 func (lb *lbBalancer) aggregateSubConnStates() connectivity.State {
 	var numConnecting uint64
@@ -298,7 +302,7 @@ func (lb *lbBalancer) aggregateSubConnStates() connectivity.State {
 			switch state {
 			case connectivity.Ready:
 				return connectivity.Ready
-			case connectivity.Connecting:
+			case connectivity.Connecting, connectivity.Idle:
 				numConnecting++
 			}
 		}

balancer/grpclb/grpclb_test.go

@@ -452,7 +452,7 @@ func (s) TestGRPCLB(t *testing.T) {
 
 	ctx, cancel = context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
-	if _, err := testC.EmptyCall(ctx, &testpb.Empty{}, grpc.WaitForReady(true)); err != nil {
+	if _, err := testC.EmptyCall(ctx, &testpb.Empty{}); err != nil {
 		t.Fatalf("%v.EmptyCall(_, _) = _, %v, want _, <nil>", testC, err)
 	}
 }