MergeVLA: Cross-Skill Model Merging Toward a Generalist Vision-Language-Action Agent