openPMD: avoid managed memory

[ax3l]

Make sure that openPMD output can run with amrex.the_arena_is_managed=0 #2009

We need to explicitly copy D2H first, e.g. here
fields:

WarpX/Source/Diagnostics/WarpXOpenPMD.cpp

Lines 1053 to 1055 in 495d052

	amrex::Real const * local_data = fab.dataPtr( icomp );
	mesh_comp.storeChunk( openPMD::shareRaw(local_data),
	chunk_offset, chunk_size );

particles:

WarpX/Source/Diagnostics/WarpXOpenPMD.cpp

Line 596 in 495d052

curr.get()[i] = aos[i].pos(currDim);

This is how plotfiles do this:
fields: https://github.com/AMReX-Codes/amrex/blob/a0790de872d6b18b563712b9f1b18323b8d00072/Src/Base/AMReX_FArrayBox.cpp#L604
particles: https://github.com/AMReX-Codes/amrex/blob/6e27136f76280e565d30ff654031d1b1ac683171/Src/Particle/AMReX_ParticleIO.H#L608

[ax3l]

• Another test: make sure that no kernels are running once we start the D2H access for I/O via GPU synchronize.
• potentially cudaMemPrefetchAsync could be called for host before I/O, but it's not clear if this only works the other way around (H2D): https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__MEMORY.html#group__CUDART__MEMORY_1ge8dc9199943d421bc8bc7f473df12e42

[ax3l]

Since we filter the particles in a temporary, pinned memory buffer (#1632): no managed memory usage there.