Also make the recursive call to a local fully typed function

Now at least as fast as tkask version

In count_and_link_tasks task counts were missed.

In make_extra_hydroloop_tasks_serial,_make_hydro_loops_dependencies was not
called (this broke the SedovBlast test as some tasks were called out of
sequence giving bad initial values).

In marktasks_mapper/serial the send/recv tasks were missing.

Conflicts:
	src/Makefile.am
	src/engine.c
	src/runner.c
	src/scheduler.c
	src/scheduler.h
	src/space.c
	src/task.c
	tests/Makefile.am

Same behaviour as old task based drift

Updated 'make check' scripts with newer tolerances for the Intel compiler and made the scripts choke if files are missing

Gravity infrastructure

Sorry, this is a bulky one. Not all of the gravity is here but most of the infrastructure is in place. It's getting tricky to maintain this separate so I'd like to merge all the innocuous changes to the master branch.
All of the gravity tasks and the logic is in place. Only the content of some of them needs to be more thoroughly checked for accuracy and I need to merge the FFT task once I am happy with it. 

With this in it should be much easier for me and Bert to re-import gizmo into the master.

Could you stress test this and verify that your usual hydro test-cases still run ? I'll bring gravity test cases in at a later point.

See merge request !212

Get the correct warning flags for the Intel compiler. Also add a -Werror for clang.

Clean up allocations and deallocations

Two improvements to the code:

 - Replaced all calls to `alloca()` by `malloc()` and `posix_memalign()` depending on use case.
 - Added cleaning up code to free up all the allocated memory. 

The second change is there to make sure we don't have memory leaks by easing the read of the memory sanitizer output. There is only one allocation that cannot be easily freed at the end of a run. In space.c (line 1327) we allocate cells in batches and then use them as a linked list. No memory leak there, just a hard unnecessary work to clean up at the end of a run.

That allows to run the EAGLE_50 example for hundreds of steps. 

See merge request !211

Correct the equation for the entropy time derivative in GADGET2_SPH

This fixes #183 meaning that energy is now correctly conserved.

It also:
 - Adds a parameter file and run script for the perturbed box test case.
 - Adds entropy to the diagnostics file.
 - Replace the MPI_AllReduce in the statistics collection by an MPI_Reduce since only rank 0 writes anyway. 
 - Gives a better documentation for the `approx_exp()` function.
 - Adds an accuracy test for the approximate maths functions (currently only exp())
 - Adds a test no make sure the symmetric and non-symmetric versions of the SPH interaction routines give the same answer. That was the origin of the energy non-conservation bug.

See merge request !210

Test 125

Added a new test to help @jwillis with vectorization. 

The old tests are now more thorough as well. 

See merge request !208