LaRa: Latents and Rays for Multi-Camera Bird's-Eye-View Semantic Segmentation