From GPU to TPU with Coding Agents: What Works, What Fails, and What It Costs?