Add clip-level video understanding for actions and changes

[tulayha]

VidXP's current visual search examines one frame per second. That works for finding visible objects and people, but a single frame cannot reliably describe movement or a change across time—for example entering versus leaving, picking something up versus putting it down, or one action happening before another.

Add an optional clip-level search path while keeping the existing frame search as the lightweight baseline.

Expected behavior:

• Split video into reusable shots or short windows with real start and end times.
• Search across multiple frames when a query depends on motion, order, or change.
• Return results through the existing evidence boards, keyframes, and downloadable clips.
• Keep clip encoding replaceable so one model is not wired throughout the application.
• Let users enable clip-level indexing only when they need it.
• Measure search quality, indexing time, memory, and storage through Add end-to-end retrieval evaluation #76.

A useful first contribution can add the shared clip representation and one measured implementation. PySceneDetect is a candidate for the initial shot boundaries; VideoPrism, PE Video, and X-CLIP are candidates for comparison rather than required dependencies.