[ICCV 2025 Oral] SceneSplat - Gaussian Splatting-based Scene Understanding with Vision-Language Pretraining