NeuroLIP: Interpretable and Fair Cross-Modal Alignment of fMRI and Phenotypic Text