Replace mmap with io_uring.

Replace mmap with io_uring.

This moves to explicit, asynchronous I/O through io_uring (Linux 5.1+),
which speeds up cold-cache behavior on rotating media by 3x or so.
It also removes any issues we might have with not fitting into 32-bit
address spaces.

If io_uring is not available, regular synchronous I/O will be used instead.
For now, there's a dependency on liburing, but it will be optional soon
for older systems.