In async write mode, we use modular arithmetic to index the output
buffer array. However, the counter variables accidently were signed,
thus resulting in negative indizes after integer overflow. That in turn
could lead to segfaults, but was depending on the memory layout of
the instance in question (which in turn depended on a number of
variables, like compile settings but also configuration). The counters
are now unsigned (as they always should have been) and so the dangling
mis-indexing does no longer happen. This bug potentially affected all
installations, even if only some may actually have seen a segfault.

I have undone a very small optimization with using pre-malloced memory,
which seems to have some issues. Now I am doing mallocs and at least in
test environment this seems to solve the issue. The code now needs more
review. If it runs flawlessly for some time, I may try to re-enable to
pre-malloc, but not necessarily: its performance benefit is very mild
(aka: I don't think it justifies introducing bigger complexities).

Some devices seem to create them and I do not see any harm in supporting that.