Decode Before You Hear: Speculative Listening for Chunk-Interleaved Full-Duplex Speech Language Models