Fix six spatial-audio rendering defects

[ctoth]

Six spatial-audio rendering fixes, one commit each with its tests. Stacked on #163 (feat/media-occlusion), which touches the same file.

What was wrong, and what changes

1. A voice-chat speaker with no known position was placed at the room origin. LiveKitSpatialAudioBridge fell back to [0, 0, 0], so a phone-call partner, or a speaker still connected from the room just left, was heard as a point at the centre of the current room at floor height. A voice is now positional only while its speaker is an entity in the current scene. Anyone else is passed as null, and MediaVoices plays that voice as a stereo source at centre pan: it has no 3D panner, so no position, no distance model and no direction. When the entity appears or leaves, the voice's sound is rebuilt in the other mode on the same track, keeping its named route. The test that asserted origin placement is rewritten to the new rule.

2. A sound's position glided across a room change. MediaService glides sound positions on its own VectorTweener, a different instance from the Client.Spatial handler's, so the handler's cancel on Scene never reached it. Scene now calls MediaService.sceneChanged(): glides in flight land on their targets at once, and the first position each sound receives afterwards is placed directly.

3. A re-Play at a moved cursor destroyed and recreated the voice. The segment comparison included start, which is the join cursor, not part of the region. Region identity is now the repeat window, loopStart..finish. On a kept voice a start seeks only when the playhead is more than MEDIA_SEEK_TOLERANCE_MS (150 ms) from the cursor. continue: false still forces a new voice.

4. Panning mode could not change on a kept voice. The mode is fixed when Cacophony creates a source, and nothing handled a later change of is3d. A Play or an Update that changes is3d for a playing key now builds the same source and region in the new mode; the old voice plays on while it loads, and the new one starts at the old one's playhead with no second fade-in. Volume, gain, pitch, spatial profile, loop count, occlusion and routing carry over.

5. Two writers moved the engine listener. ListenerPosition glided it every frame at the previously stored velocity; the listener's own EntityMove wrote the target outright when its entity glide finished, at the new velocity. A mover gets both for one step, so the listener jumped to the target and back. Both now feed one glide on the listener tween key.

6. Position glides froze in a hidden tab. The tweener schedules with requestAnimationFrame, which browsers stop for hidden tabs while audio keeps playing. It now takes a visibility source (the document by default, injectable so it stays testable without a DOM): a tween requested while hidden snaps, and glides in flight land on their targets when the document becomes hidden.

Things a reviewer should know

• Region identity (3) is not a bare field comparison. The MOO sends loopStart only for a sound that repeats. A single pass re-Played without it would never match if the client's defaulted loopStart (= start) were compared, so such a voice is kept when its region reaches back to the new cursor. A repeating sound re-Played without loopStart still gets a new voice: per the wire contract its repeat window follows start.
• Media glides finish, they are not cancelled in place (2). Cancelling would leave a sound part-way along its glide, with no further position coming if the Scene was not a room change.
• An Update's fields are now folded into the stored Play payload (4). A rebuild then replays the state in effect now, not the original Play's. This also applies to the existing moved-segment rebuild.
• Not rebuilt on an is3d change: ambisonic sounds, and preloaded sounds that were never Played.
• Segment repeat timers are not rescheduled on a kept re-Play (3). An existing test pins that behaviour.
• Not verified in a browser. The engine is mocked in these tests. The voice rebuild in (1) recreates the stream source, and with it Cacophony's muted priming element; whether Chromium leaves an audible gap at that moment needs a listen.

Checks

• npm run typecheck: clean
• npm test: 128 files, 1516 tests passed
• npm run test:audio-cache: 3 PASS lines