MLPerf Inference v6.1: pioneering agent, VLM benchmarks

First agentic workload on datacenter hardware in MLPerf, and the first model over a trillion parameters. Plus 8.85% more throughput on identical hardware since v6.0.