fix: upgrade OpenAI Realtime API to GA and resolve audio feedback loop

[PyDevDeep]

Title: fix: Upgrade OpenAI Realtime API to GA and resolve audio feedback loop

Description
This PR implements critical hotfixes to stabilize the voice streaming infrastructure. It migrates the backend from the deprecated OpenAI Beta API to the General Availability (GA) release and resolves acoustic feedback loops on the client side that were causing infinite VAD interruption cycles.

Changelog

1. Backend (OpenAI GA Migration)

• backend/services/openai_realtime.py: Removed the deprecated OpenAI-Beta header. Updated the session.update payload to match the GA schema, explicitly defining input and output audio formats.
• backend/api/ws_voice.py: Updated event listeners to consume the new GA event structures (response.output_audio.delta and response.output_transcript.done).

2. Frontend (DSP & Routing)

• frontend/src/hooks/useVoiceSession.js: Injected native WebRTC DSP parameters (echoCancellation, noiseSuppression, autoGainControl) into navigator.mediaDevices.getUserMedia to suppress acoustic feedback and prevent false-positive VAD triggers.
• frontend/vite.config.js: Fixed the dev server proxy port to route /api and /ws traffic through the Nginx reverse proxy (8080) instead of the isolated FastAPI container port (8000).
• .gitignore: Explicitly excluded frontend/node_modules/ from source control.

Acceptance Criteria Checklist
[x] OpenAI WebSocket connects using the gpt-realtime-1.5 model without 3000/4000 closure codes.
[x] Client microphone input is filtered for echo, preventing infinite response_cancelled loops.
[x] Frontend Vite proxy successfully forwards WebSocket upgrade requests to the backend.