test(agentic): add EvalScope trie benchmark protocol

[Xiangyi1996]

Summary

This draft adds a reusable EvalScope trie benchmark protocol/tooling path for agentic serving comparisons.

The protocol is based on Jiying Dong's EvalScope trie workflow and the follow-up discussion around warmup, formal sweep sizing, prompt serialization, and metric alignment. The goal is to make TRTLLM and TokenSpeed runs use the same workload, request parameters, warmup protocol, prompt serialization, and metric definitions before we compare V4-Pro agentic serving results.

What changed

• Add test/agentic_benchmark/evalscope_trie/run_evalscope_trie_sweep.sh:
  • runs two warmup passes
  • runs a formal parallel sweep
  • supports chat-completions and plain completions/tokenized-prompt mode
• Add collect_outputs.py:
  • collects EvalScope output into a CSV
  • computes Completion TPS/user
  • computes Output Token/Min/GPU
  • emits a throughput-vs-per-user-speed SVG curve
• Add README with the agreed benchmark protocol, metric definitions, and V4-Pro tokenizer/chat-template caveat.
• Add a focused unit test for the collector.

Scope

This PR does not launch or configure the server. TokenSpeed, TRTLLM, or another backend should be started separately, and backend-specific launch flags should be recorded next to each benchmark run.

Test Plan

• pre-commit run --all-files
• python -m pytest test/agentic_benchmark/test_evalscope_trie_collect_outputs.py -q
• Python compile / shell syntax checks for the new scripts

Notes

Opening as draft first so Jiying can review whether the protocol attribution and default settings match the workflow she proposed before this is marked ready for review.