fix: validate skills eval model count before running (#550)

[raywcm]

Summary

Closes #550.

benchflow skills eval currently validates the number of --model values inside SkillEvaluator.run(). When the user passes two models for one agent, the CLI has already printed the eval header and then surfaces a Rich traceback from a ValueError.

This PR moves that cardinality check to the Typer command boundary so the CLI fails before starting the eval and prints a concise validation message instead.

Changes

• Validate --model count before constructing and running the skill evaluator.
• Keep the existing accepted forms:
  • no --model
  • one --model broadcast to all agents
  • one model per --agent
• Add a CLI regression test that asserts:
  • exit code is 1
  • the friendly validation message is shown
  • no Skill eval: header is printed
  • no traceback is printed

Testing

Reproduced the current failure before the fix:

Skill eval: citation-management (1 cases)
  Agents: gemini
  Environment: docker
  Baseline skipped (--no-baseline)
...
ValueError: models length (2) must match agents (1) or be 1

After the fix:

--model may be provided once for all agents or once per --agent; got 2 models for 1 agents

Ran:

python -m pytest -q tests/test_skill_eval_cli.py tests/test_agent_cli.py
python -m compileall -q src/benchflow/cli/main.py tests/test_skill_eval_cli.py

Result:

3 passed in 2.07s