More parallelism in autorefine_triangle_soup

[sloriot]

In the function, autorefine_triangle_soup() the part of the code that is computing pairwise intersections between triangles after the box-d filtering is done sequentially. We did not bother make it parallel because in tested example it was a few percent of the total runtime. However, it seems that it can become non negligible in some cases like the one below:

./soup_autorefinement Thingi10K/raw_meshes/252786.stl 
28.9275 sec. for #2
6.2088 sec. for #3
59.6007 sec. for #1
7.53941 sec. for #4 (parallel)
#points = 2210580 and #triangles = 11549282 in 107.637 sec.
``

Check for `//TODO: PARALLEL_FOR #2` in the code.