Prototype-Aligned Latent Space Refinement in Vision-Language Models for Robust Zero-Shot Chest X-Ray Classification