Currently we calculate neighbors for vertices in batches. The idea is
to access topology data all at once, storing the result in a local node-
sized vector that can fit in CPU caches. Currently we use
`Vector<Vector<int>>` for storing the neighbors, and that works well
enough when the number of neighbors fits into the inline buffer, but
whenever it doesn't it causes many vector reallocations and scatters
the neighbor storage in arbitrary memory locations.
This commit changes to using a `GroupedSpan` to store the neighbors, in
other words contiguous storage split into groups by an array of offset
integers. This reduces local per-vertex memory usage by 20 bytes and
makes memory access more predictable.
I observed a 1.39x performance improvement for the mesh smooth filter
on a large triangulated mesh, and a 1.14x improvement for a mesh of
mostly quads. In the brush benchmark on a quad mesh the difference I
observed is a small slowdown. I'm not sure why that happens but I trust
the results from the filter a bit more.
Resolves#130797.
Pull Request: https://projects.blender.org/blender/blender/pulls/130941
583ae8990c