Wire VLM benchmarks end to end

[davidberenstein1957]

Summary

• wire benchmark registry entries to the VLM metrics they now support
• default VLM metrics to cpu plus outlines-enabled structured local generation
• replace the stub smoke script with benchmark-path end-to-end test coverage

Verification

• uv run --extra dev pytest -q tests/evaluation/test_task.py::test_vlm_metrics_fallback_to_cpu_on_auto_device tests/evaluation/test_vlm_metrics.py::test_vlm_metric_defaults_enable_structured_local_generation tests/evaluation/test_vlm_metrics.py::test_benchmark_vlm_metrics_end_to_end tests/evaluation/test_vlm_metrics.py::test_transformers_generate_routes_pydantic_response_format_to_outlines tests/evaluation/test_vlm_metrics.py::test_transformers_outlines_result_serialization tests/evaluation/test_vlm_metrics.py::test_evaluation_agent_update_stateful_metrics_with_stub_vlm