How to Scale Your Model: A Systems View of LLMs on TPUs