Revisiting Deformable Convolution for Depth Completion