perf: much faster marching cubes

[william-silversmith]

• Avoids a slow cython -> cpp point copying a numpy array to a vector
• Fast path for when all voxels in the 2x2x2 stencil match