Add optional CUDA neighbor-loop and RunSimulationCUDA

[AhmedSalih3d]

Motivation

• Provide an optional CUDA-based execution path to accelerate the per-particle interaction loops on supported GPUs.
• Allow the StillWedgeMDBC example to run on GPU when available while preserving a CPU fallback.
• Keep the CUDA integration non-intrusive and opt-in so users without CUDA are unaffected.

Description

• Add src/SPHCUDA.jl implementing a lightweight GPU path, including a CUDA neighbor-loop kernel, helpers to build neighbor cell ranges, and GPU variants of the density-diffusion and viscosity computations; export RunSimulationCUDA and CUDAAvailable.
• Wire conditional inclusion/export into src/SPHExample.jl so the package exposes RunSimulationCUDA only when CUDA.jl is present and otherwise provides CUDAAvailable() returning false.
• Update example/StillWedgeMDBC.jl to select RunSimulationCUDA when CUDAAvailable() is true, otherwise fall back to the existing RunSimulation; change example metadata to use SimpleMDBC to match the CUDA path expectations.
• Document the optional CUDA workflow in README.md and add SPHCUDA.jl to the source map; the CUDA path performs runtime checks and currently supports NoShifting/NoKernelOutput configurations and a restricted set of diffusion/viscosity models.

Testing

• Automated tests were not executed in this environment because CUDA was not available; the GPU path includes runtime checks that will error early when CUDA.functional() is false.
• Commands used during development and review include ls, rg -n ..., sed -n '...' and nl -ba src/SPHCUDA.jl | sed -n '1,240p' to inspect files and verify the new module integration.

---

Codex Task