This adds a field 'watching' in `user_for_websocket` that's True iff WATCHING,
False iff NOTWATCHING, and None otherwise (since clients don't need to know if
a user is tentative or absent). When the value of this field changes for any
user, they get added to the update buffer (like with any other change).
Removed race condition in `t_sunset_users`: `broadcast_users_update` was being
called *after* a user was removed from memory (and for each user being removed,
which was redundant). In that scenario if there's a user in the update buffer
and `t_sunset_users` wins the race between it and `t_broadcast_users_update`,
then when `t_sunset_users` calls `broadcast_users_update` a KeyError would be
raised since the user's already been removed.
Fixed unintended behaviour of `t_sunset_users`: it was removing users based on
the result of `is_visible`, so users who were actually tenative (as opposed to
absent) were being removed.
By default the buffer is exhausted every 4 seconds. This should defend against
a potential DoS against clients with JavaScript enabled. Before this, any
request with no token would generate a new user and immediately broadcast the
new user to all the websockets. It's best to lock down as much as possible the
number of places a client can cause the server to broadcasts to all the
websockets.
n9k