Fix race condition for live-polling cursors

[KyleAMathews]

Follow-up to #1826 after some more testing

There's still a slight chance that a live request is cached too long as the first long-polling request might get the same cursor as the second given networking inconsistencies, etc.

The fix is that when creating the cursor, the server should check what cursor was used by the current request and if the two cursors are the same, it should append something to ensure the cursor busts the caches.