fix: connection close should stay outside the poll loop

[garyyu]

Really don't like the macro! I just realize that try_break macro will break the poll loop, that means in case of error, we send error message to error_tx, and rely on peer::check_connection to get this error and then call stop_with_connection, and then send a message in close_channel, but at this time, unfortunately the close_channel (i.e. close_rx in conn::poll) will never be read, because poll thread already exit!

So, let's move that conn.shutdown to the outside of poll loop, to make sure an error connection is cleaned.

Another modification in this PR for peer is a small improvement to not cache the duplicated hash (for received header/block/compact_block/tx) from a peer.