Build mesh filter dataframe columns from mesh.indices

[GuySten]

Build mesh filter dataframe columns from mesh.indices

No change to any existing dataframe. Verified column-for-column against
develop for every mesh type -- regular (2D and 3D), rectilinear, cylindrical,
spherical and unstructured -- for both MeshFilter and MeshSurfaceFilter.

Problem

MeshFilter.get_pandas_dataframe reconstructs mesh element indices per axis,
including a special case for the one mesh type that is 0-based:

idx_start = 0 if isinstance(self.mesh, openmc.UnstructuredMesh) else 1

for label, dim_size in zip(self.mesh.axis_labels, self.mesh.dimension):
    filter_dict[mesh_key, label] = _repeat_and_tile(
        np.arange(idx_start, idx_start + dim_size), stride, data_size)
    stride *= dim_size

MeshSurfaceFilter.get_pandas_dataframe does the same thing again with a
different stride.

Every mesh class already exposes exactly this as indices, in bin order and
with the right base. The filter is reimplementing it, which means the index
convention for a mesh type is stated in two places, and a new mesh type needs
filter.py edited to get correct columns. It also assumes every mesh has one
index per axis with a fixed per-axis stride, which is true of the current mesh
types but is a property of those meshes rather than of meshes in general.

Changes

Both methods now take the index tuples from the mesh:

columns = [(mesh_key, label) for label in self.mesh.axis_labels]
indices = _repeat_and_tile(list(self.mesh.indices), stride, data_size)
return pd.DataFrame(indices, columns=columns)

_repeat_and_tile gains axis=0 on its np.repeat call so that an index tuple
is repeated as a row rather than being flattened.

The isinstance(self.mesh, openmc.UnstructuredMesh) special case goes away, as
does the duplicated per-axis loop in the surface filter. _mesh_current_names
is untouched.

Testing

New tests/unit_tests/test_mesh_filter_dataframe.py, parametrized over regular
3D, regular 2D, rectilinear, cylindrical and spherical, plus a separate
unstructured case:

• column names match the mesh's axis_labels, with surf appended last for the
  surface filter
• rows are exactly list(mesh.indices), in order
• structured meshes stay 1-based: first row all ones, last row equal to
  mesh.dimension
• for the surface filter, surface names cycle fastest and the element index is
  constant across one element's surface bins
• unstructured meshes stay 0-based, with a single element_index column

These pin the conventions the implementation now depends on, so a future change
to indices cannot quietly alter every user's dataframe.

I also diffed the full dataframe output between develop and this branch for
all five mesh types and both filters: columns, column order, row counts and
values are identical.

Notes

On the 0-based case specifically: UnstructuredMesh already reports

axis_labels -> ('element_index',)
indices     -> [(i,) for i in range(self.n_elements)]

so the removed idx_start branch was reproducing what indices already gave.
The old code built np.arange(0, n_elements) for the same single column, and
the two agree element for element.

This is the first of several pieces split out of #3857, and is useful
independently of that work: it removes the duplicated index logic and the mesh
type isinstance from filter.py.

Checklist

• I have performed a self-review of my own code
• I have run clang-format (version 18) on any C++ source files (if applicable)
• I have followed the style guidelines for Python source files (if applicable)
• I have made corresponding changes to the documentation (if applicable)
• I have added tests that prove my fix is effective or that my feature works (if applicable)