BGP session creation potentially caused deadlock

In AbstractBGPSessionNegotiator, handleMessage() was able to be invoked by channel.pipeline().replace(..) recursively. So when handling a BGP message, it could happen that this.state gets out of sync and exception got thrown.

In handleOpen(..) of AbstractBGPSessionNegotiator, when we new a BGPSessionImpl(), we replace the handler in channel.pipeline() to change the extend message coder. If a message arrived during the replacement, another handleMessage() will be invoked (in the same thread). However, as the BGPSessionImpl hasn't returned and session is not set, handleMessage() will throw out a BGP_FSM error.

This is causing an intermittent failure in BGPDispatcherImplTest.