CA-W3D: Leveraging Context-Aware Knowledge for Weakly Supervised Monocular 3D Detection