net/tcp: close write-worker fd on async write error (#3874)

In the IO_WATCH_WRITE handler, when the stream write handler returns
<0 (e.g. TLS async write fails after the 2s handshake timeout), the
error branch breaks out of the switch before reaching the close(s)
call that the success/pending paths hit at line 327.

The fd s is delivered to the write worker via SCM_RIGHTS and is an
independent copy in the worker's fd table — it must be explicitly
closed on every exit path.  tcpconn_release_error() signals main to
close main's copy; close(s) closes the worker's copy.  Without it
one fd leaks per failed async write, confirmed: 7 leaked fds from
51 TLS OPTIONS (handshake timeout path) vs. 0 with the fix.

Cherry-pick of bffcf00 from master.

Fixes: #3711

Author: dondetir

net/net_tcp_proc.c

@@ -312,6 +312,7 @@ inline static int handle_io(struct fd_map* fm, int idx,int event_type)
 						"handle write, err, state: %d, att: %d",
 						con->state, con->msg_attempts);
 					tcpconn_release_error(con, 1,"Write error");
+					close(s); /* we always close the socket received for writing */
 					break;
 				} else if (resp==1) {
 					sh_log(con->hist, TCP_SEND2MAIN,