Check provided MPI thread support fits requested one

There is currently a (small) problem in Chameleon in the way the call to MPI_Init_thread() is managed. You are requesting MPI_THREAD_MULTIPLE thread support, but you are not checking that MPI actually provides you this level of thread support (in case MPI cannot provide this thread support, it continues with the highest level of thread support it can manage).

This MR adds the corresponding check.

One question still remains: do you need the MPI_THREAD_MULTIPLE thread support level ? Or could MPI_THREAD_SERIALIZED be enough ?

In the testings, MPI calls are made from StarPU's dedicated thread and then from Chameleon in the check step. This requires at least MPI_THREAD_SERIALIZED, but do you need more ? It changes the way NewMadeleine (MadMPI in this case) has to be built to work properly.

(have a look here if you need a reminder )

Ping @adenis

control/control.c

@@ -96,7 +96,17 @@ int __chameleon_initpar(int ncpus, int ncudas, int nthreads_per_worker)
       MPI_Initialized( &flag );
       chamctxt->mpi_outer_init = flag;
       if ( !flag ) {
-          MPI_Init_thread( NULL, NULL, MPI_THREAD_MULTIPLE, &provided );
+          /* MPI_THREAD_SERIALIZED should be enough.
+           * In testings, only StarPU's internal thread performs
+           * communications, and *then* Chameleon performs communications in
+           * the check step. */
+          const int required = MPI_THREAD_SERIALIZED;
+          if ( MPI_Init_thread( NULL, NULL, required, &provided ) != MPI_SUCCESS) {
+             chameleon_fatal_error("CHAMELEON_Init", "MPI_Init_thread() failed");
+          }
+          if ( provided < required ) {
+             chameleon_fatal_error("CHAMELEON_Init", "MPI_Init_thread() was not able to provide the requested thread support");
+          }
       }
     }
 #  endif