Fix bug in collecting total_latencies from all threads in pgbench.

This was broken in 1bc90f7, which removed the thread-emulation. With modest
-j and -c settings the result were usually close enough that you wouldn't
notice it easily, but with a high enough thread count it would access
uninitialized memory and crash.

Per report from Andres Freund offlist.

Author: hlinnaka

src/bin/pgbench/pgbench.c

@@ -3456,8 +3456,8 @@ main(int argc, char **argv)
 		for (j = 0; j < thread->nstate; j++)
 		{
 			total_xacts += thread->state[j].cnt;
-			total_latencies += thread->state[i].txn_latencies;
-			total_sqlats += thread->state[i].txn_sqlats;
+			total_latencies += thread->state[j].txn_latencies;
+			total_sqlats += thread->state[j].txn_sqlats;
 		}
 	}
 	disconnect_all(state, nclients);