When Models Lie, We Learn: Multilingual Span-Level Hallucination Detection with PsiloQA