An interpretable, clinically-aligned AI paradigm for VTE risk prediction: an approach using LLMs and compound attention