A Large-Scale Evaluation of Speech Foundation Models