vla-eval: A Unified Evaluation Harness for Vision-Language-Action Models