Fix build with starpu1.1 which doesn't have starpu_mpi_comm_{rank,size}: let's just call RUNTIME_comm_{rank,size} which already do things properly

add cudablas library to make calls to cuda kernels (cublas here), no more calls to cublas in runtime/starpu codelets

instead of introducing RUNTIME_distributed_barrier

add cudablas library to make calls to cuda kernels (magma here, cublas will follow), no more calls to magma in runtime/starpu codelets

Also make StarPU-MPI initialization assume that in the simgrid-mpi case we do not have any outer mpi initialization

MORSE_Distributed_size, MORSE_Distributed_rank so that applications do not
hardcode the use of MPI.

Introduce RUNTIME_distributed_rank, RUNTIME_distributed_size,
RUNTIME_distributed_barrier, so that MORSE does not hardcode the use of MPI
either.

This allows to use simgrid-mpi.

Add the possibility to enter Chameleon with an existing starpu (Need to be cleanup on the starpu_initialized function)

add MORSE_Pause/Resume function to avoid CPU consumption when no tasks have to be executed - call starpu_pause/resume, do nothing for quark (no need)

move init of starpu parameters in a single place (RUNTIME_init_scheduler) and use dmdas default scheduling strategy

change the way we include our own header files --> relative to the root - when plasma is in the same env, chameleon can take some headers not belonging to it (ex: #include descriptor.h, this file states in plasma install dir also) which make compilation errors