perf: consider replacing broadcast channel with something else

A `dashmap::DashMap>` where `UserId` is the `u64` user id, `StreamId` is a `u8`. could probably be more memory efficient and faster (?) than the current solution (which is just a tokio broadcast channel + arced events). EventSender should be an unbounded sender so that we hold the lock for the least amount of time (smallmap also helps here as it is as fast as one can get with u8 keys)