Semantic-Metric Bayesian Risk Fields: Learning Robot Safety from Human Videos with a VLM Prior