DynamoLLM: Designing LLM Inference Clusters for Performance and Energy Efficiency