Use io_uring (blocked on missing copy_file_range op and liburing offload escape)

[SUPERCILEX]

If io_uring ever adds support for copy_file_range (splice doesn't do copy acceleration AFAIK), we should be able to use that to parallelize our writes.

[SUPERCILEX]

unlinkat is supported by io_uring, so we should be able to queue a getdents buffer size worth of file unlinks. Whether or not that faster, I'm not sure.

The uring command queue can be bound to a thread local.

[SUPERCILEX]

Very wip implementation for rmz: https://github.com/SUPERCILEX/fuc/tree/io_uring_rmz

[SUPERCILEX]

axboe/liburing#830