Add CUDA neighbor loop support

[AhmedSalih3d]

Motivation

• Enable the neighbor traversal hot-loop to run on CUDA GPUs for simulations with GPU-backed particle arrays.
• Provide GPU-friendly neighbor data structures by flattening per-cell neighbor lists and recording per-particle cell indices to permit contiguous GPU traversal.
• Make GPU path optional and selectable via metadata so CPU behaviour is unchanged by default.

Description

• Added FlattenNeighborCellLists! and updated exports in src/SPHNeighborList.jl to produce flattened neighbor offsets/indices for GPU traversal.
• Extended UpdateNeighbors! to optionally fill a CellListIndices array (per-particle cell index) and updated BuildNeighborCellLists! usage in src/SPHCellList.jl to call FlattenNeighborCellLists!.
• Added UseCUDA::Bool to SimulationMetaData in src/SimulationMetaDataConfiguration.jl to toggle the CUDA path.
• Introduced runtime gating (ShouldUseCUDANeighborLoop) and multiple CUDA kernel entrypoints and device kernels in src/SPHCellList.jl (several NeighborLoopPerParticleCUDA! overloads and NeighborLoopKernel* functions) to handle the different metadata/kernel/output/shifting combinations; the original threaded CPU loops remain and are used when CUDA is unavailable or disabled.
• Refactored per-interaction callsites to pass a small ParticleFields view and a SimMetaDataType type for more uniform kernel/call signatures required by the CUDA wrappers.

Testing

• No automated tests were executed for this change (no test run requested).
• Commands executed while preparing this PR included repository inspection and edits such as ls, rg, sed, multiple apply_patch actions, and git commit -m "Add CUDA neighbor loop support".

---

Codex Task