feat(voice): add realtime voice sessions

[big-saltedfish]

What changed

• Add a voice entry below the composer that starts a dedicated realtime voice session view.
• Keep the existing text conversation and history views isolated from voice-session rendering.
• Display realtime transcripts and task execution progress, with tool details collapsed by default.
• Support barge-in by stopping assistant playback when the user starts speaking.
• Restore voice-session history through metadata.conversation_id and mark sessions with metadata.source = voice-gateway.
• Proxy realtime WebSocket traffic through the local server so browser clients do not need custom authentication headers.

Why

This aligns the quickstart voice experience with the Cloud Agents web console while keeping the existing chat flow unchanged.

Security and deployment

• Restrict the local voice proxy to loopback origins.
• Use short-lived, one-time connection keys and normalize WebSocket close codes.
• Disable the voice entry when the deployment does not advertise voice capability, including the hosted Vercel demo.
• No credentials, private keys, local environment files, or personal paths are included.

Validation

• npm run test --workspace=client — 145 tests passed.
• npm run test --workspace=server — 6 tests passed.
• npm run lint --workspace=client — no errors (2 pre-existing warnings).
• npm run build --workspace=client — passed.
• npm run build --workspace=server — passed.

[vercel]

@big-saltedfish is attempting to deploy a commit to the Murphy's projects Team on Vercel.

A member of the Team first needs to authorize it.