feat: add video frame retrieval

[raoabinav]

Evaluator Summary

• Abi added frame-level video retrieval by sampling local video frames, embedding them with CLIP, and indexing each sampled frame as a searchable LEANN passage.
• Design choice: each frame gets a stable video-frame-<hash> ID plus timestamp, frame index, frame size, duration, source path, media type, and model metadata, so search results point back into the original video.
• The implementation builds from precomputed CLIP frame arrays and disables query-time recomputation for frame vectors, matching the image-retrieval design while keeping visual retrieval separate from audio transcription.
• Quality bar: 9 targeted tests mock decoding and model inference to avoid codec/model downloads while covering sampling, metadata, stable IDs, and retrieval wiring, plus lint, format, type, lockfile, and whitespace checks.

[yichuan-w]

I don't think CLIP is a very good embedding model to use nowadays, especially for video, maybe Qwen3VL embedding is better

[raoabinav]

Parking this PR for now to keep the review queue focused on #347. I’ll revisit/reopen this work later as a smaller clean PR once the content-hash PR lands.