Be more conservative when splitting self-task into subs. We don't want to trigger a rebuild straight after.

Modified the splitting rule for self tasks to be less conservative. Added proper documentation of the splitting conditions.

Also create functions to decide when to rebuild and when to split a task. Update the rebuild and split condition to correctly use kernel_gamma.

Add new function to decide on recursion with in a sub_pair task. Makes sure all recursion tests are identical.

Completes d8025de1

Use dx_max_sort_old and not dx_max_sort when checking whether cells have been sorted in DOPAIR1_BRANCH()

Updated the tests to use the new cell properties for sorting and not the old ones. All tests now run smoothly. Make sure the xparts array actually exist in a cell before reinitialising its particles.

Do not call the difffloat.py script with the additional deprecated parameters formerly used for rho_dh

Do not call the difffloat.py script with the additional deprecated parameters formerly used for rho_dh

Test periodic bc

This tests the periodic boundary conditions which are generated in `space_getsid()` and used in `runner_dopair1_density`.

An 8x8x8 mesh is generated that is periodic and 26 pair interactions are computed for the following cells:

* All 8 corner cells
* The cell at the centre of each face
* The cell at the centre of each edge

These checks are performed for both unperturbed and perturbed cells.

See merge request !350

Output the position of the cell in the output of the periodic boundary condition test. Better tolerance requirements.

Corrected mistake in the periodic boundary test that prevented compilation when optimisations are switched off.

Make the difffloat script use a new line in the files to set a limit below which values are ignored as too close to zero.

 - Use the correct value of 'h_max' in the construction of the caches in the vectorized DOPAIR1 routine. (fix #312)
 - Update the test27 unit test to also randomly perturb the smoothing length values.
 - Set a correct value for h_max in the test27 case and not just a loose upper-bound.
 - Call the non-vectorized version of DOPAIR1 for the corner interactions.

See merge request !348

EAGLE_25 now produces the same number of time-steps whether running with or without vectorization.