VLX-Go: Robust Embodied Tracking through Multi-Stage Learning and On-Policy Refinement