Nice project. One thing I would focus on is allowing for a zero litter work flow. In Hazelcast I wrote a similar system, and if properly used, it is litter free.
I would like also to have benchmarks between the current Java API, and you library. Otherwise it is not obvious, does your library outperform standard implementation, or not.
Just like fio, I supported the possibility to run the benchmark not with io_uring, but with the standard API. But, these results are quite difficult to interpret. Let's say, with the same number of threads io_uring produces more IOPS, but the load on the CPU is higher.