Evaluating the Ability of Large Language Models to Generate Verifiable Specifications in VeriFast