docs: Voxtra as adapter only, LiveKit role, and browser-based calling

[byamasu-patrick]

Follow-up to #1. These three commits were pushed after #1 had already been merged, so they did not land on main. This PR brings them in.

Changes to the integration roadmap

• Voxtra is a telephony adapter only. The AI agent implementation (STT, LLM, TTS, knowledge base, tools, agent logic) stays in the luso8 ai-engine; no new code lands in the Voxtra repo. The runtime piece to build is a luso8 ai-engine Asterisk worker that uses the Voxtra adapter and reuses the existing voice pipeline (transport swaps from LiveKit to Asterisk, not the brain). Reflected in the decisions, diagram, principles, integration seams, Phase 1, env map, and risks.
• Where LiveKit fits. Clarifies that Voxtra plus Asterisk replaces the carrier/PSTN and AI legs, but the browser still needs WebRTC.
• Browser-based human calling. Agents dial out and listen in or jump in from the browser, so a WebRTC path is a first-class requirement. Documents the options: keep LiveKit for the browser leg near term; target a browser-to-Asterisk WebRTC path (SIP.js/JsSIP over WSS, chan_pjsip + rtpengine); Kamailio terminates WebRTC at scale.

Documentation only, no code or infra change. No em-dashes per house style.