Set up sorting boxes on the CuPy backend

[max-models]

Closes #726

What changed

• SortingBoxes._set_boxes (pic/sorting.py): n_cols is computed with math.sqrt (host bookkeeping). _neighbours is built on the host with the pyccel initialize_neighbours and moved to the active backend once with xp.to_cunumpy (setup data).
• Particles setup with boxes (pic/base.py):
  • n_boxes uses math.prod instead of xp.prod(xp.array(...)).
  • is_domain_boundary holds Python bools read from a host copy of domain_array.
  • _get_neighbouring_proc (SPH only) reads a host copy of domain_array. Before this, the pyccel distance got 0-d device arrays and ParticlesSPH setup failed on CuPy.
• put_particles_in_boxes / do_sort: assign_box_to_each_particle, assign_particles_to_boxes and sort_boxed_particles are wrapped in cunumpy.kernels.PyccelKernel with declared outputs. On NumPy they run the kernel directly, as before. On CuPy this is an explicit host fallback: every call copies device → host → device.

Time loop of non-SPH models

For a non-SPH model (e.g. VlasovAmpereOneSpecies), setting boxes_per_dim alone does not sort anything during the time loop. Pushers only call put_particles_in_boxes when sorting_boxes.communicate is true, which is SPH only. Box sorting runs only:

• during setup, via draw_markers(sort=True), when SortingParameters(do_sort=True);
• every sort_step steps in sim.py, via do_sort(), when EnvironmentOptions.sort_step > 0.

Both paths now work on CuPy through the host fallback above. The fallback still has to be ported (#694).

Remaining for #694

• Device-side box sorting: CUDA versions of the three sorting kernels, to replace the PyccelKernel host round trip.
• _check_and_assign_particles_to_boxes still compares a device max_in_box in a Python if, which forces an implicit sync.
• SPH evaluation and ghost-box communication on the device. I checked that ParticlesSPH setup works on CuPy. Drawing markers with a domain needs CUDA launches, which the fake CuPy in cunumpy 0.6.1 cannot run, so nothing after setup was exercised for SPH.

Tests

New check_sorting_boxes_on_cupy in pic/tests/test_sorting.py:

• creates Particles6D(boxes_per_dim=(4, 3, 2)) on NumPy and on CuPy, and checks that neighbours, boxes, next_index and cumul_next_index match;
• runs do_sort() on the same markers on both backends and checks that the markers and box arrays match;
• creates ParticlesSPH with boxes and checks that is_domain_boundary (Python bools), the boundary boxes, the neighbouring procs and neighbours match NumPy.

It runs in two tests:

• test_sorting_boxes_on_cupy uses the real CuPy and is skipped without a GPU.
• test_sorting_boxes_on_cupy_fake_cupy uses cunumpy's fake CuPy in a serial subprocess with CUNUMPY_FAKE_CUPY=1, on rank 0 only. It follows the pattern of test_cuda_args_domain_fake_cupy.

Results:

• Before the fix, the fake-CuPy test fails with TypeError: type ndarray doesn't define __round__ method. With only the setup fix, it fails with argument must be numpy.ndarray in the sorting kernels.
• pytest src/struphy/pic/tests/test_sorting.py: 53 passed, 3 skipped (MPI-only tests and the GPU test).
• mpirun -n 2 pytest --with-mpi src/struphy/pic/tests/test_sorting.py: 55 passed, 1 skipped (the GPU test).

Merge order

Based on devel. It does not depend on the unmerged CUDA PRs: on devel, Particles can already be created on the fake CuPy backend.

🤖 Generated with Claude Code