Fix to #363. Drifting decision needs to use the same multipole information as the actual tree walk. Also fix some gravity_cache padding values.

When padding the gravity cache, use -2 * cell->width for the particle position to make sure this is an invalid (but reasonable) position.

Split the P2P and M2P into two separate (vectorizable) loops in the leaf-leaf gravity P-P interactions.

Do the P-P gravity calculations in the frame of the cells for more accuracy in the distance calculations.