Data race in Kokkos OpenMP initialization

[PGXavier]

Singletons are initialized in a parallel region in Kokkos_OpenMPExec.cpp, line 145+. A critical region fixes that. (Found with Intel Thread Inspector, Intel 15. Debugged with Christian Trott.)

 #pragma omp parallel
    {
      const int rank_rev = m_map_rank[ omp_get_thread_num() ];
      const int rank     = pool_size - ( rank_rev + 1 );
#pragma omp critical
      {
          m_pool.at(rank_rev) = HostSpace::allocate_and_track( "openmp_scratch", alloc_size );
      }
      new ( m_pool[ rank_rev ] ) OpenMPexec( rank , ALLOC_EXEC , reduce_size , thread_size );
    }
/* END #pragma omp parallel */

[crtrott]

Should be done now.