bench(fast_io): IOCP vs stdio write throughput on Windows (#1899)

[oferchen]

Summary

• Adds crates/fast_io/benches/iocp_vs_stdio.rs, a Criterion bench that pits the Windows IOCP writer path (IocpDiskBatch) against std::fs and BufWriter<File> baselines.
• Four cells per payload size (4 KiB, 64 KiB, 1 MiB) for 12 measurement points: iocp_default, iocp_concurrent_ops_8, stdfs, bufwriter_64k. Each cell writes 1000 files into a fresh tempfile::TempDir per iteration.
• Closes the "wired but never measured" gap from Ensure release flow builds distributable artifacts #1868 with repeatable data on whether the IOCP path actually helps under typical and high-concurrency loads (task Fix RPM artifact paths and cargo xtask alias #1899).

Implementation notes

• Bench is cfg-gated to target_os = "windows" + the iocp feature. On Linux and macOS the file compiles to a stub fn main() that prints a skip line, so cargo bench -p fast_io stays cheap on every other host.
• IocpDiskBatch is the only writer surface that consumes IocpConfig::concurrent_ops (via submit_write_batch in crates/fast_io/src/iocp/disk_batch.rs). The bench drives it through the same begin_file -> write_data -> commit_file sequence the receiver disk-commit thread uses.
• bufwriter_64k matches the default IocpConfig::buffer_size so the comparison isolates overlapped submission from buffering.
• Payloads are seeded by a deterministic LCG to defeat NTFS sparse-region folding without introducing a random-number crate dependency.
• See the file's module doc for the run recipe and the outcome -> action mapping that feeds back into Fix RPM artifact paths and cargo xtask alias #1899 and Cache x86 checksum SIMD feature detection #1929/Optimize vectored rolling checksum aggregation #1930.

Test plan

• cargo bench -p fast_io --bench iocp_vs_stdio on a Windows host produces three throughput rows per iocp_default/iocp_concurrent_ops_8/stdfs/bufwriter_64k cell.
• cargo build -p fast_io --bench iocp_vs_stdio on Linux and macOS hosts compiles to a no-op stub that runs and prints the skip line.
• CI fmt+clippy passes on every platform matrix.