feat: video support in huddle calls

**Motivation**
Teams that have moved their standups and working sessions into Buzz huddles (our team runs all its calls there now) still have to leave the platform whenever a call needs faces or screen-sharing — pairing, demos, design reviews, or just seeing teammates. Audio-only huddles cover a lot, but video is the remaining reason to keep a Zoom/Meet dependency next to an otherwise self-hosted, self-sovereign workspace.

**Proposed solution**
Add video (camera, and ideally screen-share) to huddles. Roughly in the spirit of the existing design:

- Extend the huddle transport beyond the current Opus-over-WebSocket relay (`src/audio/` in buzz-relay) with a video track — either a new frame type on the same socket or a parallel `/huddle/{channel_id}/video` endpoint, keeping the relay forwarding opaque frames between authenticated, membership-checked peers as it does today.
- Reserve/define the corresponding Nostr event kinds for video track lifecycle, consistent with how huddle join/leave/end events already work.
- Bandwidth is the hard part: pure fan-out relay of video frames scales worse than audio, so this may justify simulcast or an optional SFU mode for larger rooms, even though audio deliberately avoids one.

Even a v1 limited to small rooms (2–5 participants, single resolution, no screen-share) would remove the external-tool dependency for most day-to-day calls.

**Alternatives considered**
- Keep using Zoom/Meet/Jitsi alongside Buzz — works, but splits identity, membership, and (eventually) recording/transcription across systems, and defeats the self-hosted premise.
- Embed/link an external self-hosted Jitsi or LiveKit per channel — possible today with no Buzz changes, but participants lose NIP-42-authenticated membership gating and the signed huddle event trail.
- WebRTC P2P mesh negotiated over Nostr events, bypassing the relay — avoids relay bandwidth but breaks the "all reads and writes flow through the relay" architecture and behaves badly behind NAT without TURN.

**Additional context**
- ARCHITECTURE.md "Huddle Audio — WebSocket Opus Relay": https://github.com/block/buzz/blob/main/ARCHITECTURE.md — notes recording and per-track publishing are planned with kinds reserved; video tracks could slot into that same track model.
- Duplicate search: none found — open "video" issues (#4143, #5193, #5415, etc.) all concern video *file uploads/validation*, not real-time calls.

(Filed from my account with the help of my Buzz agent, which drafted this after digging through the architecture docs.)