feat(dataviewer): support multiple camera angles in episode viewer

[akzaidi]

Component

data-management/viewer/ (frontend + backend)

Problem Statement

The current dataviewer episode viewer renders a single camera stream per episode. Bimanual VLA policies (TwinVLA, π₀, RDT-1B) require up to 3 camera views during training and inference:

• Front/ego-centric camera — workspace overview
• Left wrist camera — left arm end-effector view
• Right wrist camera — right arm end-effector view

RoboTwin 2.0 datasets include front_image, wrist_image_left, and wrist_image_right per episode step. LeRobot datasets may include additional camera keys. Without multi-camera visualization, operators cannot:

1. Verify camera coverage and alignment across views
2. Inspect occlusion or lighting issues in individual camera streams
3. Correlate spatial relationships between arm-mounted and workspace cameras
4. Quality-check the exact visual inputs the VLA model receives during training

Proposed Solution

Add multi-camera display support to the episode viewer:

1. Auto-detect camera keys from LeRobot dataset metadata (scan observation.images.* keys)
2. Grid layout — display all camera views simultaneously in a responsive grid (1×1 for single camera, 1×3 for three cameras, 2×2 for four, etc.)
3. Camera selector — allow toggling individual cameras on/off via a camera panel
4. Synchronized scrubbing — all camera views stay frame-synchronized when scrubbing the timeline
5. Per-camera zoom — click a camera view to expand it to full width while keeping others visible as thumbnails
6. Camera labels — display the LeRobot image key name as an overlay on each view

Technical Notes

• LeRobot v3.0 stores images as MP4 videos under videos/{camera_key}_episode_{idx}.mp4
• Camera keys are defined in dataset info.json under features.observation.images
• The backend already streams video frames; the change is primarily frontend layout and state management
• Consider using CSS grid with auto-fit for responsive layout

Acceptance Criteria

• Episode viewer detects and displays all available camera streams from the dataset
• Camera views are frame-synchronized with the timeline scrubber
• Individual cameras can be toggled on/off
• Clicking a camera view expands it; clicking again returns to grid
• Camera key names displayed as overlay labels
• Works with single-camera datasets (no regression)
• Works with 3-camera bimanual datasets (RoboTwin 2.0, TwinVLA Tabletop-Sim)

Context

• VLA training pipeline: training/vla/ (branch feat/vla-twinvla-robotwin)
• Bimanual robot types: evaluation/sil/bimanual_robot_types.py
• Tracking issue: feat(dataviewer): build & lint remediation — tracking issue #115 (dataviewer build & lint remediation)
• Related: test(dataviewer): TypeScript frontend testing infrastructure and coverage — tracking issue #210 (frontend testing tracking issue)