GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks