Compare realistic agentic inference performance

Long-context, multi-turn inference performance on Google TPUv7 Ironwood, OpenAI Jalapeño, MI355X, GB300 NVL72, GB200 NVL72, B200, H200, H100 and RTX Pro. TPUv8i, SambaNova SN50, Rubin NVL72 and MI455X UALoE72 are next.