Efficient 3D Semantic Scene Completion with Sparsity-Aware Attention, Adaptive Feature Selection and Depth-Aware BEV Decoder