Precision Meets Efficiency: Optimizing Multimodal MRI Brain Tumor Segmentation via Lightweight Encoders and Model Pruning