Add optional CUDA neighbor-list and GPU neighbor-loop utilities

[AhmedSalih3d]

Motivation

• Provide a GPU-capable neighbor-list and neighbor-loop implementation so neighbor search and per-particle interaction loops can be offloaded to CUDA, with focus on NeighborLoop and neighbor-list utilities.
• Keep the existing CPU path intact and enable incremental adoption by loading CUDA helpers only when CUDA.jl is available.

Description

• Added src/SPHCUDANeighborList.jl which defines CUDACellGrid, CUDANeighborList, GPU buffer allocation, and CUDA kernels for computing per-particle CellIds, CellCounts, ParticleRanges, ParticleOrder, precomputing neighbor-cell indices, and the device-side neighbor loop.
• Implemented high-level CUDA helpers: AllocateCUDANeighborList, ConstructNeighborOffsets, BuildNeighborCellListsCUDA!, UpdateNeighborsCUDA!, and NeighborLoopCUDA! to mirror the CPU neighbor-list workflow on the GPU.
• Updated src/SPHExample.jl to conditionally include and re-export the CUDA utilities only when CUDA.jl is present (HasCUDA), preserving default CPU behaviour.
• Documented usage and installation notes in README.md with a minimal sketch showing CUDACellGrid allocation, UpdateNeighborsCUDA!, and NeighborLoopCUDA!.

Testing

• No automated tests were run for this change.

---

Codex Task