Support configurable TCP socket options (TCP_USER_TIMEOUT)

[Karrq]

Problem

When using TcpTls (or plain Tcp) under network conditions with packet loss, TCP's default retransmission behavior causes unbounded delivery delays. The default TCP_RETRIES2 (15 retries) means TCP will keep retransmitting for up to ~13-30 minutes before giving up, while the connection appears healthy from the application's perspective.

Currently, the only socket option set is TCP_NODELAY (in both tcp/mod.rs and tcp_tls/mod.rs connect() and poll_accept()). There's no way for consumers to configure retransmission behavior.

This is problematic for applications that implement their own reliability/recovery layer on top of msg-rs - TCP's silent retransmissions become redundant and introduce unbounded latency that the application can't control or detect.

Proposed solution

Add socket-level configuration to the Tcp and TcpTls client/server configs. At minimum:

• TCP_USER_TIMEOUT (Linux) - max time (ms) unacknowledged data can remain before TCP aborts the connection. This is the most impactful option for bounding retransmission delays.

These would be applied in connect() and poll_accept() right alongside the existing set_nodelay(true) call:

let stream = TcpStream::connect(addr).await?;
stream.set_nodelay(true)?;

#[cfg(target_os = "linux")]
if let Some(timeout) = config.user_timeout {
    // setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, timeout_ms)
}

Nice-to-haves (follow-up)

• SO_KEEPALIVE + TCP_KEEPIDLE / TCP_KEEPINTVL for faster dead connection detection
• Exposing these through the Control channel for runtime tuning (note: SubDriver doesn't currently wire up control_rx, so this would also need work in msg-socket)