[Feature] Improve LingBot WebRTC streaming controls and lifecycle

[lzx1413]

Description

Improve LingBot streaming reliability and interactive WebRTC control behavior.

The PR introduces the session-aware actor scheduler for LingBot streaming, native RTP frame delivery with bounded latest-frame buffering, stronger WebRTC session cleanup, and synchronized camera-control snapshots. It also refreshes the v1/v2 examples and stream-server documentation.

Motivation

Long-running LingBot sessions need bounded memory, isolated per-session state, and predictable cleanup. The previous control path could retain stale directional input, while generated-frame metadata and the HUD could diverge from the controls actually applied to a chunk.

This change also removes an unnecessary JPEG/base64 round trip for bidirectional LingBot output and makes session progress and queue behavior observable.

Type of Change

• Bug fix
• New feature
• Performance improvement
• Code refactoring
• Documentation update

Changes Made

• Add a session-aware actor streaming scheduler, health barriers, resource permits, and scheduler validation coverage.
• Preserve LingBot KV-cache cursors across actor/FSDP boundaries.
• Route bidirectional output as native frames into RTP, with bounded latest-frame queues that drop stale frames under client backpressure.
• Add WebRTC DataChannel-open timeout, disconnect grace handling, bounded pipeline-close calls, and session lifecycle metrics.
• Expose runtime progress, latency, queue high-watermark, dropped-frame, and scheduler telemetry through the DataChannel and demo UI.
• Replace stale directional-command buffering with latest control-state snapshots.
• Keep the per-chunk HUD and applied_controls metadata aligned with the controls submitted to the model.
• Separate translation (WASD / arrows) from rotation (IJKL), add reset-pose behavior, and set realtime translation scale to 3.0 for both v1 and v2 examples.
• Update English and Chinese stream-server, scheduler, TURN, and LingBot v2 documentation.

Testing

• Targeted unit tests pass
• WebRTC integration tests pass
• Manual stream-server/TURN startup and health-check validation performed
• Benchmarks/validation tooling added
• New regression tests added
• Documentation updated

Test commands:

.venv/bin/python -m pytest -q \
  tests/unit/orchestrator/test_parallel_worker_stage_actor.py \
  tests/unit/orchestrator/test_streaming_pipeline_orchestrator.py \
  tests/unit/pipelines/lingbot_world_fast/test_control_alignment.py \
  tests/unit/pipelines/lingbot_world_fast/test_runtime_baseline.py \
  tests/unit/pipelines/lingbot_world_fast/test_service_action_loop.py \
  tests/unit/pipelines/lingbot_world_fast/test_stream_example.py \
  tests/unit/service/test_service_routes.py \
  tests/unit/service/test_webrtc_session_manager.py

.venv/bin/python -m pytest -q \
  tests/integration/test_webrtc_api.py::TestWebRTCBidirectional

git diff --check

## Checklist

- [x] Code follows project formatting and lint rules
- [x] Commit hooks pass
- [ ] Full pytest tests/ suite run
- [x] New tests added for new functionality
- [x] Documentation updated
- [x] Commit history is split by scheduler, transport, lifecycle, controls, and docs

## Related Issues

N/A

## GPU Architecture Support

No kernels were added or modified.

## Performance Impact

Bidirectional LingBot output now bypasses generated-frame JPEG/base64 serialization and uses bounded latest-frame buffering. Under a slow client, stale frames are intentionally dropped to preserve freshness and bound memory use.