inference.py: native video preprocessing (no qwen-vl-utils)

inference.py: embed_video (base's own video path; purely additive)

from_pretrained: revision pinning

Upload inference.py with huggingface_hub