Fix IndexError in _pic_swarm_in_mesh on MPI ranks with empty cell partition

[hardik-corintis]

Description

Fixes #5077.

In _pic_swarm_in_mesh, parent_cell_nums_local contains -1 sentinels for points not
owned by the current rank. On ranks with an empty cell partition, cell_closure has shape
(0, k) and NumPy rejects -1 as an axis-0 index, causing an IndexError when
interpolating from a Submesh onto its parent mesh in parallel.

Fix: index cell_closure only on visible_idxs (the non-sentinel entries), pre-filling
the rest with -1. The same issue exists in the extruded branch: _parent_extrusion_numbering
floor-divides parent_cell_nums_local, and -1 // N == -1 in Python, so sentinels pass
through into base_parent_cell_nums and cause the same crash. Both branches are fixed.

Two regression tests added in tests/firedrake/submesh/test_submesh_interpolate.py using
8 MPI ranks:

• test_submesh_interpolate_3Dcell_2Dfacet_empty_rank_8_processes: interpolates from a
  Submesh of the x=1 face of UnitCubeMesh(2, 2, 2) onto the parent mesh.
• test_submesh_interpolate_3Dcell_extruded_empty_rank_8_processes: interpolates from
  ExtrudedMesh(UnitSquareMesh(1, 1), layers=3) (6 cells) onto a UnitCubeMesh(2, 2, 2),
  guaranteeing at least two ranks own zero extruded cells.

[leo-collins]

I've just had a look at this now (on the main branch). I don't get the IndexError you described. However, it does fail the assertion at the bottom, but this is because subm is an immersed manifold. We handle these by expanding the bounding boxes into an N-hypercube, where N is the geometric dimension of the mesh, and points not on x=1 were being found and evaluated on subm because they are within the default tolerance of the mesh. The test passes after setting subm.tolerance=0.1.

I'm also not sure that the reasoning for your fix is correct; see my comment here #5077

[leo-collins]

Actually, if I run this on 8 ranks I do get the IndexError you described. The tests in test_swarm.py I mentioned don't actually input any points onto the rank which has zero cells, which is what is happening here. In that case I think this fix is good.

[leo-collins]

I think this is good. The same thing probably happens for the extruded case too - can you add a test for this (and a fix if necessary)?

[leo-collins]

Actually, if I run this on 8 ranks I do get the IndexError you described. The tests in test_swarm.py I mentioned don't actually input any points onto the rank which has zero cells, which is what is happening here. In that case I think this fix is good.

[hardik-corintis]

Thanks @leo-collins for the quick review. I have applied all the inline suggestions and added the test and fix for the extruded meshes.