Vision-language models for zero- and few-shot industrial anomaly detection: an evidence-grounded survey