SHANKS: Simultaneous Hearing and Thinking for Spoken Language Models