This is on purpose. Assume that the query result has a well-defined order.
If I want rows 21 through 30 from a multi-partition table, those rows could
theoretically happen to be rows 21 through 30 from some partition
-- maybe a partition that happens to contain all of the early rows -- or it may be rows 1 through 10 from some partition
(all earlier rows coming from other partitions).
The only safe assumption is that no individual partition has to produce more than its first "limit + offset" total rows, so long as it produces those rows in the correct order.