Make Scatterer "X"-aware

[IgorBaratta]

• Add option to use MPI_Isend/MPI_Irecv :
  • It is not necessary to use host buffers and intermediate memory transfers with CUDA-aware MPI (if the right allocator is used).
  • Neighbor collectives are not supported (yet?) in CUDA-aware MPI .
• Add allocator for indices so pack/unpack kernels can be implemented on the targeted device.

Example (SYCL):

using DeviceScatterer = common::Scatterer<sycl::usm_allocator<std::int32_t, sycl::usm::alloc::shared>>;

auto queue = select_gpu_queue(comm);
sycl::usm_allocator<std::int32_t, sycl::usm::alloc::shared> allocator(*queue);
DeviceScatterer sct(map, bs, allocator);