Created a horizontal maximum with intrinsics, which turns out to be slower than a simple for loop over the vector in AVX at least.

Pass vectors by address instead of value and use 2 vectors to find neighbours in runner_doself2_force_2.

runner_doself2_force_vec_3 now passes addresses of the interaction cache to the interaction function.

Created a runner_doself2_force_vec function. Multiple variations, runner_doself2_force_vec vectorises straight from the serial. runner_doself2_force_vec_2 vectorises by caching properties needed in the interaction. runner_doself2_force_vec_3 builds on runner_doself2_force_vec_2 and stores the interactions in a secondary cache and performs interaction after looping over all possible pj particles.

Created functions to evaluate only W and dWdx separately along with vectorised versions. Also removed unneeded FMAs when kernel constants are zero.

Updated kernel_deval_1_vec and kernel_deval_2_vec to compute both branches of the Cubic Spline kernel and combine the result with masks.

Gravity multi dt

Some improvements to the gravity code:

 - Corrected typo in 4th order vector powers.
 - Use a user-defined opening angle for the distance checks.
 - Increase accuracy of the exact gravity calculation.
 - Only do exact gravity calculation if a matching result file is not present in the same directory.
 - Reinstate the dumping of individual task timers.
 - Added script to plot the accuracy of the calculation.
 - Crash if the M2L kernel is called with a length smaller than the softening length.

See merge request !331

Use short int consistently for unlocks. Check that the waits and unlocks don't overflow.

Two small changes:
 - `task->nr_unlock_tasks` is a `short int` so we should be consistent with the types in `scheduler_set_unlocks()`.
 - When debugging, check that we don't overflow the number of `wait` or `unlocks`.

As I am about to introduce some gravity tasks that will have a lot of unlocks I would like to be safe first.

See merge request !330

Only loop over particles that are within range and active. Removed if statement that should never be true.

Improved the way of computing the exact accelerations by removing the need of recalculating exact forces that already exist

Use the multipole acceptance criterion to build the long-range interactions and not just the geometry of the mesh.