feat(tagger): Qwen-VL NL captions, pose extraction (YOLOv8-n), slide-panel model selector

feat(tagger): pose extraction + Qwen3-VL optional NL captioner + WD14x3 ensemble weights recalibration