Refactor GroupCall participant management

This refactoring brings a number of improvements to GroupCall, which I've unfortunately had to combine into a single commit due to coupling:

- Moves the expiration timestamp field on call membership state to be per-device
- Makes the participants of a group call visible without having to enter the call yourself
- Enables users to join group calls from multiple devices
- Identifies active speakers by their call feed, rather than just their user ID
- Plays nicely with clients that can be in multiple calls in a room at once
- Fixes a memory leak caused by the call retry loop never stopping
- Changes GroupCall to update its state synchronously, and write back to room state asynchronously
  - This was already sort of halfway being done, but now we'd be committing to it
  - Generally improves the robustness of the state machine
  - It means that group call joins will appear instant, in a sense

For many reasons, this is a breaking change.

spec/test-utils/webrtc.ts

@@ -431,6 +431,7 @@ export class MockCallMatrixClient extends TypedEventEmitter<EmittedEvents, Emitt
 export class MockCallFeed {
     constructor(
         public userId: string,
+        public deviceId: string | undefined,
         public stream: MockMediaStream,
     ) {}