feat(web): add one-shot camera photo input

[yimi528]

Summary

• Add an explicit Take photo control to the WebUI composer.
• Add user-controlled continuous camera observation for Omni Realtime models.
• Capture bounded JPEG frames at approximately 1 FPS and retain at most the latest 8 frames in memory.
• Forward observation frames through input_image_buffer.append without creating response.create.
• Stop camera tracks on cancel, stop, page hide, disconnect, mute, sleep, and component unmount.
• Add Gateway observation events, model capability discovery, translations, styles, documentation, and tests.

Scope

This PR supports:

• One-shot photo capture as a normal WebUI attachment.
• Explicit continuous camera observation after the user clicks “Continuous observation”.
• JPEG capture up to 1280×720, with Base64 payloads capped at 256 KiB.
• Omni Realtime transport only; legacy Audio models and native video remain unsupported.

Continuous observation does not proactively produce model responses. Frames become part of the Realtime visual context for subsequent user interaction.

Validation

• npm run lint
• npm test --workspace web — 106 tests passed
• Observation runtime tests — 6 tests passed
• Gateway contract tests — 13 tests passed
• npm run build --workspace web
• git diff --check

Privacy and compatibility

• Camera permission is requested only after the user presses Take photo.
• Continuous observation requires explicit user opt-in.
• Raw camera frames and Base64 data are kept only in bounded in-memory buffers.
• Frames are not persisted and are not forwarded to the backend ACP Agent.
• Camera capture stops when the page is hidden, the connection is lost, or the user stops observation.