When Agents Slow Down: Understanding LLM Agents’ Test-Time Strategies via Elo-per-token Analysis