unread_data: Ensure deterministic ordering of unread message rows.

Earlier, in `get_raw_unread_data` the ordering was applied inside
the CTE.

Once we leave the CTE scope and do a join, SQL makes no promise
about preserving the row order unless we re-specify ORDER BY in
the outer query.

Since, there was no ORDER BY clause in the outer query it was
resulting in a random ordering of the entries. This bug was caught
by `test_unreads_case_insensitive_topics` failing in a flaky way.

This commit fixes the bug.

(cherry picked from commit ccc82976dc)

zerver/lib/message.py

@@ -857,6 +857,7 @@ def get_raw_unread_data(
             "flags",
             "recipient_id",
         )
+        # Descending order, so truncation keeps the latest unreads.
         .order_by("-message_id")
     )
 
@@ -909,9 +910,13 @@ def get_raw_unread_data(
                     "recipient__type",
                     "recipient__type_id",
                 )
+                # Output in ascending order. We can't just reverse,
+                # since the CTE join does not guarantee that it
+                # preserves the original descending order.
+                .order_by("message_id")
             )
 
-            rows = list(reversed(user_msgs))
+            rows = list(user_msgs)
         finally:
             cursor.execute("SET enable_bitmapscan TO on")
         return extract_unread_data_from_um_rows(rows, user_profile)