SPAD: Specialized Prefill and Decode Hardware for Disaggregated LLM Inference