SocialNav-SUB: Benchmarking VLMs for Scene Understanding in Social Robot Navigation