feat: VoiceCoach v2.1 - GPU local, Ollama LLM, settings panel, token server separation

- pipeline.py: switched LLM from OpenClaw/Claude to local Ollama, added settings loading, sentence aggregator, async SpeechBrain, tuned VAD
- main.py: simplified (no more token server thread, just pipeline)
- token_server.py: NEW - standalone token + settings API server
- docker-compose.yml: all services on GPU machine, added token-server service, CUDA speaches, Ollama LLM
- web-client/index.html: added settings panel with sliders
- web-client/app.js: settings management, localStorage persistence, push to server
- web-client/style.css: settings panel styles
- web-client/nginx.conf: proxy to token-server instead of pipecat-agent
- patch_livekit.py: updated with proper indentation-safe patching

pipecat-agent/Dockerfile

@@ -6,6 +6,11 @@ RUN apt-get update && apt-get install -y libsndfile1 ffmpeg && \
 WORKDIR /app
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
+
+# Patch pipecat LiveKit transport bugs
+COPY patch_livekit.py /tmp/patch_livekit.py
+RUN python3 /tmp/patch_livekit.py
+
 COPY . .
 
 HEALTHCHECK --interval=30s --timeout=10s --start-period=30s \