Attention Satisfies: A Constraint-Satisfaction Lens on Factual Errors of Language Models