Self Pre-training with Topology- and Spatiality-aware Masked Autoencoders for 3D Medical Image Segmentation