Demystifying Platform Requirements for Diverse LLM Inference Use Cases