Io_uring based IO for evicted liquid arrays

Currenlty we issue blocking read/write to disk: https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/cache/mod.rs#L163
https://github.com/XiangpengHao/liquid-cache/blob/main/src/liquid_parquet/src/cache/mod.rs#L443

This blocks the entire runtime. We want to explore how io_uring can help performance.

Prior art: https://github.com/tokio-rs/tokio-uring
https://github.com/tokio-rs/tokio/discussions/4703

We want a benchmark where x-axis is thread count, y axis is io-usage.

To measure disk usage: https://docs.rs/sysinfo/latest/sysinfo/struct.DiskUsage.html
Also want to plot the latency.
To vary the thread count: https://github.com/XiangpengHao/liquid-cache/blob/main/benchmark/clickbench/clickbench_client.rs#L166

=====

We want to compare:

1. current approach
2. tokio::fs::File
3. your io_uring implementation. Blocking io_uring; thread pool based io_uring; and non-blocking io_uring.

=====

Run clickbench, hits.parquet ~15GB.
Run query 20: https://github.com/XiangpengHao/liquid-cache/blob/main/benchmark/clickbench/queries/queries.sql#L21

https://github.com/XiangpengHao/liquid-cache/blob/main/benchmark/clickbench/clickbench_client.rs#L134

Paper references:
https://vldb.org/pvldb/vol16/p2090-haas.pdf

Start with the blocking reads and writes in src/liquid_parquet/src/cache/mod.rs and thread-count handling in benchmark/clickbench/clickbench_client.rs. Run ClickBench query 20 against hits.parquet, then compare the current approach, tokio::fs::File, and the requested io_uring variants. Done means benchmarks varying thread count with disk usage and latency plots.