embed_video: native preprocessing (frame tensors and torchcodec path decode); drops the qwen-vl-utils dependency; bitwise-equal to the previous implementation and the raw base on identical inputs

Add embed_video: the frozen base's own video path (qwen-vl-utils preprocessing, official defaults), gate-guarded like images; bitwise-verified against the base and the reference loader

Remote code: batched embed_text_batch / embed_audio_batch entry points

AutoModel + trust_remote_code support (remote-code loading; existing loaders unchanged)