Do different partition and dataset scenarios that trigger carry situations

…f packets in wrapup

Make sure we catch near-range stragglers if there is unfinished work. That way we
"survive" massive scrambling within the range of one tile.

Adapt test cases accordingly; also test dups which is trivial.

Use a "virtual corrected frame ID" and correct wrap already during decoding.
Wrapping is detected directly after processing packets and is corrected as early as possible.

To not get mixed up in "old" and "new" epoch, the decoded header is stored
in the carry buffers.

Test for wrapping was flaky for 128 scrambling,
test it only for less aggressive scrambling.

In reality, inconsistencies in the data stream are rare,
and the packets are almost always in order, so jump
detection will work with high reliablity.

Also fix comments from before

The tests are still flaky sometimes, more work to be done

* Account for stragglers in test
* Refine test matrix to exclude combinations that can "trip"
  the packet reordering. The test is already much harder than the actual data stream,
  which is almost always in order and complete with only very few
  missing, scrambled or duplicate packets.