feat(livekit): mix loading-indicator into the single published audio track

[tigranbs]

Summary

Publishes the loading-indicator audio as part of the single agent audio track, mixed in server-side, instead of on a separate loading-audio track.

LiveKit is an SFU — it forwards each published track independently and never mixes audio server-side. Many subscribers only ever play one audio track: custom browser clients that attach a single <audio> element, and SIP bridges that down-mix to one phone stream. A dedicated second track therefore never reached them. This change mixes the loading clip into the one published tts-audio track so every single-track subscriber hears it.

How it works

A single audio "pump" task is the sole writer to the published NativeAudioSource:

• Idle / speaking: TTS PCM is captured pass-through — byte-identical output to before.
• Indicator active: the pump emits continuous 10 ms frames from the looping clip and sums in any buffered TTS (saturating i16 add — the same approach as LiveKit's own AudioMixer), with a short fade-out on stop.

Both modes live in one task, so there is never a second writer racing the source (two writers on one NativeAudioSource interleave frames rather than mixing).

Fixes included

• No truncated speech. The TTS buffer now back-pressures the producer instead of dropping the oldest samples, so rapid multi-sentence speak output plays in full (previously the middle could be cut off).
• No crash on non-standard sample rates. The source queue depth is computed as a valid non-zero multiple of 10 ms for any sample rate (e.g. 44.1 kHz no longer trips the libwebrtc assertion).
• Client-supplied loading clips of any supported rate/channel layout are resampled once at load time (adds rubato) to the published track format.
• Reconnect / disconnect / teardown simplified to the single source + pump.

Compatibility

• Replaces the dedicated loading-audio track; an audio session now publishes exactly one audio track (tts-audio) whether or not a loading clip is configured.
• WebSocket API unchanged: the loading_audio config and loading_start / loading_stop commands behave the same; clients that don't use them are unaffected.

Testing

• cargo fmt and cargo clippy --all-targets --all-features -- -D warnings clean.
• Full unit/integration suite green, including new resampler and mixer unit tests; the libwebrtc-native tests pass via the dedicated isolated CI step.
• Docs (docs/websocket.md, docs/api-reference.md, docs/livekit_integration.md) updated to the single-track model.