fix(fast_io): prevent io_uring SEND deadlock under TCP backpressure (#1872)

[oferchen]

Summary

• Fixes docs: document canonical source repository #1872. An IORING_OP_SEND on a back-pressured TCP socket would sit in the kernel until the send buffer drained, holding submit_and_wait and starving any concurrent RECV completion side, producing an apparent deadlock during daemon-mode bidirectional transfers.
• Adds a PollAdd(POLLOUT) readiness gate (with a linked timeout) in front of every submit_send_batch flush. SEND SQEs only enter the kernel once the socket reports writable, so submit_and_wait is bounded by readiness rather than peer drain. This mirrors upstream rsync's select()-based bidirectional I/O loop in io.c:perform_io.
• Transient EAGAIN/EWOULDBLOCK SEND CQEs and ETIME/ECANCELED poll CQEs re-arm the readiness wait without surfacing fatal errors.

Why poll-add gating instead of an interleaved peek loop

Each IoUringSocketWriter/IoUringSocketReader already owns a private ring (so the SEND ring never carries RECV SQEs). The cleanest containment for the deadlock is therefore inside the writer's own submit_send_batch: defer the SEND submission until the kernel signals room. This adds zero coupling between the writer and reader rings and keeps the existing API surface intact, whereas an interleaved peek/timeout loop would require sharing CQE handling between two rings that today have no shared state.

Test plan

• CI: cargo nextest run -p fast_io --all-features -E 'test(io_uring) or test(socket) or test(send)' (Linux must exercise test_socket_send_no_deadlock_under_backpressure_1872; macOS/Windows skip via #[cfg(target_os = "linux")]).
• CI: full nextest matrix (Linux musl, Windows, macOS) on stable.
• CI: fmt + clippy + interop workflows.

The new regression test:

1. Opens a loopback TCP pair and shrinks SO_SNDBUF/SO_RCVBUF to 4 KiB.
2. Pre-fills the writer's kernel send buffer with a sentinel byte until a non-blocking write returns EAGAIN, then restores blocking mode.
3. Spawns separate writer and drain threads, each backed by its own io_uring ring (IoUringPolicy::Enabled).
4. Asserts the writer reports the full 64 KiB payload within a 20 s wall-clock deadline (without the fix this would loop indefinitely).
5. Skips gracefully when io_uring is unavailable.