Add an internal limit to the batch-finishing queue so it does not grow without bound

It is possible for records to fail faster than they are promised, if you are (contrived) spin loop producing failing records. This should push memory issues back to your application, rather than OOMing within the client.