OmniVLA-RL: A Vision-Language-Action Model with Spatial Understanding and Online RL