Compress neighbor cell adjacency storage

[AhmedSalih3d]

Motivation

• Reduce allocations and pointer-chasing from Vector{Vector{Int}} per-cell neighbor lists by using a compact adjacency layout.
• The stencil size is bounded (max 3^D - 1), so a fixed-size dense layout plus per-cell counts can improve cache locality in the inner neighbor loops.
• Preserve existing neighbor-iteration ergonomics so calling code in hot loops requires minimal changes.

Description

• Add CompressedNeighborCellLists (dense Matrix{Int} of shape (max_neighbors, n_cells) plus Counts::Vector{Int}) to store adjacency compactly.
• Provide NeighborCellIndices iterator and Base.getindex so for NeighborIdx in NeighborCellLists[cell] remains valid.
• Implement EnsureNeighborCellCapacity! and update BuildNeighborCellLists! to populate the compressed layout and return the structure.
• Update alloc/rebuild sites in src/SPHCellList.jl to allocate CompressedNeighborCellLists(...) and accept the returned value from BuildNeighborCellLists!, and re-export CompressedNeighborCellLists from the top-level module.

Testing

• Ran git diff --check which reported no whitespace/errors.
• Ran julia --project=. -e 'using SPHExample' which precompiled and loaded the package successfully.
• Ran a focused neighbor-list unit check with julia --project=. -e 'using Test; using SPHExample.SPHNeighborList; ... BuildNeighborCellLists!(...) ; @test collect(lists[2]) == [3,4]' which validated neighbor enumeration and succeeded.
• Ran the full test suite with julia --project=. -e 'using Pkg; Pkg.test()', which errored in an existing time-stepping test due to a mismatched Δt method signature unrelated to the neighbor-list changes.

---

Codex Task