Add YAML scenario profiles and recursive evaluation aggregation

[ShuxinLin]

Summary

This extracts the benchmark and evaluation work from #487 while intentionally excluding every OpenAI-agent-related change.

Benchmark scenario selection

• replace the category-specific text lists with all.yaml and lite.yaml profiles
• add named selectors such as fcc_lite, fcc+fmsr_all, lite, and all
• retain support for custom YAML profiles and plain-text scenario lists
• update the scenario-suite documentation and unit tests

Evaluation

• recursively discover trajectory JSON files in nested directories
• optionally filter scenarios using the benchmark selectors
• group score summaries by runner and model
• write a single aggregate report for recursively collected results
• update evaluation tests for recursive loading, filtering, and aggregation

Scope

The PR changes 22 files, limited to:

• benchmarks/
• src/benchmark/
• src/evaluation/

It excludes OpenAI-agent code, CLI options, launcher scripts, workspace configuration, documentation, examples, and test fixtures.

Testing

uv run pytest src/benchmark/tests/test_scenario_suite_runner.py src/evaluation/tests

Result: 123 passed.