Does RLVR Extend Reasoning Boundaries? Investigating Capability Expansion in Vision-Language Models