Argus — FFai's vision-language component

Named for Argus Panoptes, the all-seeing hundred-eyed watchman: image captioning, visual Q&A, and video understanding.

Backend plan (Phase 4): mistral.rs — the inference engine built on candle (FFai's tensor spine, so buffers are shared without conversion) — running Qwen-VL / LLaVA-class models with quantization. Video understanding composes ffai-media::sample_frames (rff-backed keyframe sampling) with per-frame or windowed captioning into a timed track.