Spec1: backbone-agnostic extraction dispatcher + InternVideo2/VideoMAEv2 loaders

[jinmang2]

Part of #16 (Spec 1 — Foundation).

Make feature extraction work for any backbone, inference-only.

Scope

• scripts/extract_features.py → --backbone {i3d,clip,videomae,videomaev2,internvideo2} dispatcher via build_extractor(); cache *_<backbone>.npy; keep segment_features (dim-agnostic). Legacy I3D path byte-identical.
• New src/features/internvideo2.py, src/features/videomaev2.py — OpenGVLab native weights (not transformers-native) → custom loader (prefer validated HF export). dim from model config.
• Reuse existing dataset-keyed layout + unified manifest.

Done when

• Each backbone produces cached features of the documented dim.
• I3D path unchanged (regression).

[jinmang2]

Dispatcher landed (commit on PR #22): --backbone routes to any registered extractor via src/features/extract.py (run_extraction/save_path_for, unit-tested); import decord made lazy; legacy I3D path intact.

Deferred to a GPU session (needs OpenGVLab weights + GPU): InternVideo2 and VideoMAEv2 custom loaders. Keeping this issue open for those.

[jinmang2]

Consolidated into PR #22 (tracked in the PR checklist) to keep management to PR + Epic only. Code for done items is in PR #22; remaining items live in the PR's Spec-1 checklist.