Add new let and duplicate starpu  implicit algorithm in GroupTree repository

uniform kernel.

Before you'd have to split the total amount of particles between the
number of node ...

export tarball. Add check for time in python script. Small change in the
timer in the tests (not that much).

That way no need of using a file to exchange them. Split the mpi job in
two. Change a little bit informations printed in the job.

So we get average time of executing and test to check if the result is
good.

Also add the sbatch script that run tests.

The implicit part reproducing explicit mapping works well.
But on the explicit side, it turn out that if all morton indexes ain't
used, it segfault.
Gotta investigate it.

testBlockedImplicitAlgorithm and doesn't require a mapping file anymore.

To do it properly, MPI_Init is done in the main and MPI_Finalize in the
algorithm destructor.

of explicit mpi.

It works well when the number of leaves is dividable by the number of
mpi process.
Else there is some offset.

of the octree.

dag comparing.

Still there is errors in the result of the core algorithm.

Add a few more information to the ones simgrid use.
Put the same taskname in the explicit mpi.
Add source code to compare dag.
Add script to compile and run tests.

can't remember.

the changes are the new data layout, the deprecation of the POD cell
adaptors.

Changed tests are:
  - testBlockedAlgorithm,
  - testBlockedAlgorithmDyn,
  - testBlockedMpiAlgorithm,
  - testBlockedChebyshev,
  - testBlockedMpiChebyshev,
  - testBlockedRotation,
  - testBlockedRotationCompare,
  - testBlockedRotationMpi,
  - testBlockedSpeedOperators,
  - testBlockedTaylor,
  - testBlockedTree,
  - testBlockedUniform,
  - testBlockedUniformBench,
  - testBlockedUniformCompare

When using starpu number of threads should be given by env variables (and not command line parameter)