Graceful shutdown support

[tanyaofei]

I noticed that there is currently a graceful shutdown feature, but I found that if a request is still being processed and takes a long time, the client may receive a ConnectionCloseException. I think it should behave more like Tomcat’s graceful shutdown: first stop accepting new requests, then wait for all ongoing requests to complete before actually shutting down.

1. Close discovery.
2. Wait for all requests to finish processing.
3. Shutdown the transport
4. ....

[artem-v]

This behavior can be conditional, by default we close immediately.