tls_openssl: fix openssl_tls_async_connect

The `fd` value here is used with `poll()`/`select()` in the event that `SSL_connect()`
gives an `SSL_ERROR_WANT_WRITE` error. So we need to be polling/selecting on the fd that
is used by the underlying socket, which is the fd passed as an argument to the function,
rather than `con->fd`.

It seems to me that `handle_io()` updates `con->fd` to match the received fd number only
when the socket wants reading and not when it wants writing. When the socket wants writing
the correct fd to use is the one that is passed in the function argument, not the one
that is in `con->fd`.

The reasons this mostly worked before are:

1.  on a box that is not heavily-loaded you often end up with the same worker setting up
    the TLS session as opened the socket in the first place, so the fd in `con->fd happens
    to be correct.
2.  if `SSL_connect()` happily connects straight away, without an `SSL_ERROR_WANT_WRITE`,
    then you never run into a code path that uses the value in `fd` anyway.

Author: Your Name

modules/tls_openssl/openssl_conn_ops.c

@@ -657,7 +657,6 @@ int openssl_tls_async_connect(struct tcp_connection *con, int fd,
 	poll_err=0;
 	elapsed = 0;
 	to = timeout*1000;
-	fd = con->fd;
 
 #if defined(HAVE_SELECT) && defined(BLOCKING_USE_SELECT)
 	FD_ZERO(&orig_set);