Experiences from Benchmarking Vision-Language-Action Models for Robotic Manipulation