Context-Aware Mixture-of-Experts Inference on CXL-Enabled GPU-NDP Systems