With this change, views now read every reactive value through `useBehavior`, as
`architecture.md` already prescribes. Where a view reads a plain stream, the view
model now exposes a `Behavior`.
A tile shows "Waiting for media" for as long as its MatrixRTC member
cannot be matched to a LiveKit participant. Log each transition of that
match and of the tile's waitingForMedia state so rageshakes can tie a
stuck tile to the LiveKit participant and track events.
It turns out that the timers which repeatedly poll the RTP stats of each video track all run independently of each other and can add up to a small but constant sink of CPU. Meanwhile we can replace these timers with HTMLVideoElement 'resize' event listeners, which is way more efficient and reacts instantly to orientation changes.
The rule of thumb to avoid resource leaks is that you should never call ObservableScope methods in a callback unless the ObservableScope is directly passed to or created inside that callback. I had a go at codifying this as a lint rule.
I found our code's internal model of ringing a little overgrown (it had superfluous states like 'unknown') and difficult to extend with metadata or callbacks relating to ring attempts. By modeling ringing instead as a stream of ring attempts, where each attempt has an intent, a recipient, and an eventual outcome (accept/decline/timeout), I find it more natural to work with.
This makes room for a future 'try again' callback to allow ringing someone again after a timeout, and also forced me to look for a simpler solution to the duplicate leave sound effects. I exposed the intent of the ringing attempt to the call UI so I can later use it in the header.
