docs: expand LiveKit Agents analysis with deep code insights

Add detailed analysis of latency optimization techniques:
- Preemptive generation: trigger LLM before VAD confirms
- Adaptive video FPS: 1fps speaking, 0.3fps silent
- AEC warmup: 3s protection against false interruption
- False interruption recovery with 2s silence window
- Transcript buffering during agent speech
- TTFB/TTFT tracking for end-to-end latency breakdown
- Three-layer interruption detection architecture

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Author: claude