Use io_uring on Linux where available (>= kernel 5.1)

[ned14]

Apparently can do 1.7M 4k IOPS per CPU core. As PCIe maxes out at a theoretical 2M 4k IOPS due to PCIe bus latency, that's pretty great for O_DIRECT i/o. It may even be worth deploying this for the synchronous implementation on Linux, though synchronous reads can reach 4M 4k IOPS/core on Linux.

https://lwn.net/Articles/776703/

Example of use: http://git.kernel.dk/cgit/fio/plain/t/io_uring.c

[ned14]

This actually should be folded into a new nonblocking_file_handle which implements non-blocking synchronous i/o, and coroutinised resumable i/o. This would allow us to drop async_file_handle.