Investigating Decoder-only Large Language Models for Speech-to-text Translation