This is motivated by the fact that if you have a configuration
where there is a large number of boxes intersection (such as
k*n^2, with n being the number of faces), then the size
of the concurrent container can be very large and thus
the memory footprint would be huge.
Instead, since the box_intersection_d call is now parallel,
we can just do callbacks in the tree directly. The only thing is,
we want to have these (heavy) callback calls to be roughly balanced.
For this, we random shuffle the range of faces.
Same runtime as before, no more huge memory footprint!
Mael Rouxel-Labbé
7af0704cea