stderr from a bzr+ssh smart server process should not be sent to the client

The warnings should not be lost, but rather logged to a place which is
available to the codehost maintainers (eg rsync'd away from the
production server or something similar.)

On the whole I think it does make sense to display them on the client, because sometimes they're the best clue as to what is going wrong. However, they should be distinguished from local messages, maybe by adding a prefix to each line. However, we can't use ssh stderr for messages the user must see, because there's no equivalent when operating over tcp or http.