bench(fast_io): IOCP vs io_uring under matched workloads (#1868)

[oferchen]

Summary

• Add crates/fast_io/benches/iocp_vs_iouring_matched.rs: a matched-workload bench that drives IocpDiskBatch on Windows and IoUringDiskBatch on Linux through identical payload sizes (4 KiB / 64 KiB / 1 MiB), file counts (1000), per-iteration temp-dir lifecycles, and a deterministic 4 KiB tiled payload pattern.
• Every host (Linux, Windows, macOS) runs a portable std_baseline cell that uses File::create + write_all so per-host throughput can be normalised before any cross-host comparison. The two kernels cannot be compared head-to-head; the normalised ratio is the cross-platform signal.
• Register iocp_vs_iouring_matched in crates/fast_io/Cargo.toml with harness = false. The bench file compiles on every platform; non-target-OS bodies fall back to the std_baseline-only group so Criterion's harness = false contract holds everywhere.

Cells per platform

• Linux (all(target_os = "linux", feature = "io_uring")): iouring_default, iouring_concurrent_ops_8 (sq_entries = 8), iouring_sqpoll (sqpoll = true), std_baseline. The io_uring cells are env-gated via OC_RSYNC_BENCH_IOURING_RING=1; the SQPOLL cell additionally requires OC_RSYNC_BENCH_IOURING_SQPOLL=1, matching the existing iouring_sqpoll_vs_regular.rs recipe.
• Windows (all(target_os = "windows", feature = "iocp")): iocp_default, iocp_concurrent_ops_8 (concurrent_ops = 8), std_baseline.
• Every other host: std_baseline only.

Normalisation approach

Divide each platform-specific cell's throughput by that host's std_baseline throughput for the same payload size, then compare the resulting ratios across the Linux and Windows runs. The ratio strips out host-specific storage stack effects so the residual delta reflects the kernel-async dispatch style itself. This is documented in the bench file's module doc with a worked example.

Test plan

• CI fmt+clippy passes.
• CI nextest (stable) passes on Linux / macOS / Windows.
• Linux runner with io_uring opt-in: OC_RSYNC_BENCH_IOURING_RING=1 cargo bench -p fast_io --bench iocp_vs_iouring_matched produces all four Linux cells.
• Windows runner: cargo bench -p fast_io --bench iocp_vs_iouring_matched produces all three Windows cells.
• macOS runner: bench compiles and emits only the std_baseline cells.