rpc: pass errors from async_accept and avoid leaks #4142

According to the [boost::asio documentation](http://www.boost.org/doc/libs/1_55_0/doc/html/boost_asio/reference/basic_socket_acceptor/async_accept/overload2.html),
the function signature of the handler must be:

    void handler(
      const boost::system::error_code& error // Result of operation.
    );

We were binding *all* the arguments, instead of all but the error,
resulting in nullary function that never got the error. Fix this
by adding an input argument substitution.

Make sure conn object always gets cleaned up by using a
`boost::shared_ptr`.

This makes valgrind happy - before this commit, one connection object
always leaked at shutdown, as well as can avoid other leaks, when
for example an exception happens.

Also add an explicit Close() to the !ClientAllowed path to make it similar
to the normal path (I'm not sure whether it is needed, but it
can't hurt).

That option hasn't existed for a long time.

Fixes #4156.

The problem is that the boost::asio::io_service destructor
waits for the acceptors to finish (on windows, and boost 1.55).

Fix this by keeping track of the acceptors and cancelling them before
stopping the event loops.