Flash-Fusion: Enabling Expressive, Low-Latency Queries on IoT Sensor Streams with LLMs