mesa/glthread: decrease the batch size for better perf scaling