If you still have the full strace (or can generate a new one), can you check how many times io_setup is invoked during the process run? It’s extremely unlikely, but it might be that the process itself is eating up all available AIO slots due to some bug.
Otherwise, it might be a kernel bug? That’d be quite odd too, since we never saw this, and I believe this part of kernel code is relatively stable.
Also, can you confirm that this crash happens consistently all the time you try to start the lxd daemon? And that io_setup fails every time?