Improve agent workload simulation and metrics

[jamesdborin]

Summary

• tag agent requests with a persistent per-agent UUID by default, with opt-out and deterministic prefix support
• treat model output as opaque assistant state and append a valid synthetic tool-call/result pair
• add --ignore-eos for exact output-length workloads
• report completed-agent end-to-end p50, p90, and p99 latency in console and CSV output
• document the updated agent-mode behavior

Why

Agent-mode benchmarks need stable session affinity without assuming that random model output contains valid tool calls. They also need exact output-length generation and full-agent lifetime percentiles to characterize stateful multi-turn workloads accurately.

Validation

• cargo fmt --check
• cargo test (18 tests passed)
• cargo clippy --all-targets -- -D warnings attempted; blocked by two pre-existing warnings in untouched rust/src/cli.rs (large_enum_variant and manual_unwrap_or)