Scene-Agnostic Traversability Labeling and Estimation via a Multimodal Self-supervised Framework