Buffered cell_split matthieu

Here is an alternative version to what you wrote.

Instead of sorting the buffers and the particles at each recursion level of space_split_recursive() (via cell_split()), I only sort the buffers and at the end, when all the recursion is done, I re-arrange the part/gpart array using the indices that I store in the cell_buff structure. That means that each part is only accessed twice. Once to initially construct the buffers and once to do the final swap. In @nnrw56's version the particles are swapped once per level (and 3 times per level in the old master). That also allows to do the part<->gpart linking only once and not once per level.

Pedro, what do you think of this ? There might be some room for micro-level improvement in the function cell_reorder_parts() using similar tricks to your memswap function.

Note, this is built on top of your branch "cell_split".