MR-CLIP: Efficient Metadata-Guided Learning of MRI Contrast Representations