feat(fast_io): IOCP completion-port pump for batched writes (#1897)

[oferchen]

Closes #1897.

Summary

• Add CompletionPump in crates/fast_io/src/iocp/pump.rs: a shared
  completion-port driver that spawns a worker thread looping on
  GetQueuedCompletionStatusEx and dispatches each entry to a per-op
  CompletionHandler keyed by the address of its OVERLAPPED. This
  is the Windows analogue of an io_uring CQE drain.
• Wire the pump module into crates/fast_io/src/iocp/mod.rs and
  re-export CompletionHandler, CompletionPump, IocpPumpConfig,
  oneshot_handler, and (Windows-only) iocp_post_completion from
  crates/fast_io/src/lib.rs.
• Mirror the surface in crates/fast_io/src/iocp_stub.rs with
  Unsupported-returning stubs so the crate still compiles on Linux
  and macOS.
• Clean shutdown via PostQueuedCompletionStatus with a sentinel
  completion key; Drop runs the same path so abandoned pumps still
  reap their worker thread.
• Status interpretation: NTSTATUS read out of OVERLAPPED.Internal,
  translated to Win32 DOS errors for the most common overlapped-I/O
  outcomes (STATUS_END_OF_FILE, STATUS_CANCELLED,
  STATUS_INSUFFICIENT_RESOURCES, STATUS_IO_TIMEOUT); anything
  unrecognised passes through io::Error::from_raw_os_error.
• All unsafe stays inside fast_io per the workspace unsafe-code
  policy. No new unsafe in transfer, core, cli, or engine.

Why this PR is foundational

#1898 (batched-write API matching the IoUringDiskBatch surface),
#1929/#1930/#1928 (concurrent IOCP paths in the receiver / disk-commit /
delta-apply hot paths), #1931/#1932 (stress tests under load), #1900
(CI matrix expansion), and #1899 (benchmark harness) all depend on
having a proactor in place. Today every IocpReader::read_at and
IocpWriter::write_at allocates a private port and inline-calls
GetQueuedCompletionStatus; the audit at
docs/audits/windows-iocp-benchmark.md (sections 1.1, 1.2, 5.1)
documents the gap. Subsequent PRs will layer batched submission on
top of this pump.

Audit findings (pre-implementation)

• crates/fast_io/src/iocp/completion_port.rs:17-86 - existing
  CompletionPort is a simple RAII wrapper, only associate(handle, key)
  and handle(). No drain function.
• crates/fast_io/src/iocp/file_reader.rs:131-149 and
  crates/fast_io/src/iocp/file_writer.rs:140-156 - both inline-call
  GetQueuedCompletionStatus per op, blocking the calling thread.
• docs/audits/windows-iocp-benchmark.md:136-141 - "each IocpReader /
  IocpWriter allocates its own CompletionPort ... no shared port,
  no thread-pool drainer, no per-process proactor."
• No prior PR has shipped a pump. PR feat: add Windows IOCP async file I/O in fast_io crate #3299 added the per-handle scaffold.
  PRs found: feat: add Windows IOCP async file I/O in fast_io crate #3299 (scaffold), docs(fast_io): propose Windows IOCP benchmark plan (#1899) #3510 (benchmark plan Fix RPM artifact paths and cargo xtask alias #1899). No pump.

What is left for downstream PRs

• Fix packaging asset paths and legacy binaries #1898 - batched-write API. Layer IocpDiskBatch on top of the
  pump, mirroring IoUringDiskBatch (crates/fast_io/src/io_uring/disk_batch.rs).
• Cache x86 checksum SIMD feature detection #1929, Optimize vectored rolling checksum aggregation #1930, Enable SIMD rolling checksums on x86 #1928 - concurrent IOCP paths. Wire the pump into
  the receiver basis-read path, the disk-commit thread, and the
  delta-apply hot path. Replace per-op GetQueuedCompletionStatus
  calls in IocpReader::read_at / IocpWriter::write_at with pump
  registration.
• Accelerate bandwidth parser ASCII scanning #1931, Optimize sparse zero detection #1932 - stress tests for batched-write throughput and
  cancellation under load.
• Make legacy wrappers optional and tighten packaging checks #1900 - extend the Windows CI matrix to exercise the IOCP path.
• Fix RPM artifact paths and cargo xtask alias #1899 - benchmark harness (audit already merged in docs(fast_io): propose Windows IOCP benchmark plan (#1899) #3510).

Test plan

• cargo fmt --all -- --check clean
• cargo clippy -p fast_io --all-targets --all-features --no-deps -- -D warnings clean
• cargo clippy --workspace --all-targets --all-features --no-deps -- -D warnings clean
• cargo nextest run -p fast_io --all-features -E 'test(iocp)' (23 tests pass on macOS stub path including 3 new pump-stub tests)
• CI: fmt+clippy, nextest (stable), Windows (stable; runs the real pump dispatch test), macOS (stable), Linux musl (stable), interop
• Windows CI exercises pump_dispatches_overlapped_write_completion (writes via WriteFile against FILE_FLAG_OVERLAPPED handle, asserts the pump's worker fires the registered handler with the correct byte count)
• Windows CI exercises pump_post_completion_round_trip (manual PostQueuedCompletionStatus with synthetic OVERLAPPED, asserts dispatch)