sTriangulation.neighbour_simplices() method is expensive

[jimc101]

From some tests I have done, the time taken to identify neighbour simplices in a spherical triangulation scales with N^{2}, where N is is the number of simplices. This seems to be consistent with the underlying Fortran code, which includes two nested loops over the element set. Have you thought about ways in which the efficiency of this operation could be improved; for example, by exploiting aspects of the geometry or by using threading to perform the task in parallel?