Though, with the large dumps it seems to me that with every slight change in the actual data the entire dump gets downloaded again. I'm already using uncompressed dumps in the hope that the output is more stable and RSYNC better able to recognize unchanged parts. But I guess that most changes in the dumped data simply result in all subsequent data being that misplaced compared to what RSYNC reads against, that it's like downloading the whole file again in the end.
For me the rsync differential algorithm on dump files just works with no special preparations. I think you need to investigate this from the rsync side, not PostgreSQL. For example, if you change a dump file just by adding or deleting one character near the beginning, how does rsync respond to that?