Support trace-replay datasets such as AIPerf AgentX in presets

[r4victor]

Hasn't verified yet.

Presets could only benchmark datasets as independent requests, so a dataset of recorded agent sessions (e.g. the InferenceX AgentX traces) could not be used to optimize for real multi-turn traffic.

• Add an aiperf profile row to the agent prompt's dataset options, with --public-dataset and the inferencex-agentx-mvp scenario of SemiAnalysis's AIPerf fork.
• Require recorded multi-turn sessions to be replayed with their turn order and timing, with the same replay options in every trial and a shorter duration allowed for trials than for the final benchmark.
• Allow a replay to lose requests up to the tool's own failure threshold instead of requiring every request to succeed.
• For a replay, compute output_tok_per_s over the replay duration rather than the tool's total run time, and take per_user_tok_per_s from the tool's per-user metric, since idle sessions keep far fewer than concurrency requests in flight.