You don't have to wait for requests to complete, so slow or broken connections don't cause hangs. IIRC, zmap[1] takes this approach. This allows the "probe" stage to be executed essentially as quickly as you can send packets, while you lazy load the results
Nmap is already asynchronous, as you described multiple requests are multiplexed (cf select(2), epoll(7), kqueue/kevent depending on your platform...)
The core difference between nmap and zmap is that the latter is stateless. Which allows a much higher throughput, but also prevents from applying some smart detection mechanisms. Thus, both tools are complementary, and no, there is no work being done to make nmap stateless.
