The Ultra-Scale Playbook: Training LLMs on GPU Clusters