Natural Language Supervision for General-Purpose Audio Representations