On the Evaluation of Speech Foundation Models for Spoken Language Understanding