3D hierarchical dual-attention fully convolutional networks with hybrid losses for diverse glioma segmentation