Inline half-step in neighbor loop

[AhmedSalih3d]

Motivation

• Integrate the half-step state update into the neighbor interaction pass so per-particle half-step updates (acceleration, position, velocity, density) are computed inline with neighbor interactions and avoid a separate HalfTimeStep pass.
• Ensure gravity is applied consistently during the half-step accumulation and make the neighbor loop responsible for producing Positionₙ⁺, Velocityₙ⁺, and ρₙ⁺ for the next staged pass.

Description

• Added NeighborLoopPerParticleHalfStep! variants that mirror existing NeighborLoopPerParticle! signatures (handling NoKernel/KernelOutput and Shifting modes) which accumulate dρdt, Acceleration, kernel/shift outputs where applicable, compute AccelerationMax, apply gravity with ConstructGravitySVector, and write Positionₙ⁺, Velocityₙ⁺, and ρₙ⁺ inline using the provided dt₂.
• Replaced the first neighbor-loop + separate HalfTimeStep sequence in the main simulation loop with a single call to NeighborLoopPerParticleHalfStep! and updated the @timeit labels to reflect the consolidated step ordering.
• Left the original NeighborLoopPerParticle! implementations intact for the second neighbor pass and preserved dispatch patterns for compatibility across kernel/shifting/MDBC modes.

Testing

• Commands executed during development: ls, rg -n "ComputeInteractionsPerParticleNoKernel" src, several sed -n inspections of src/TimeStepping.jl and src/SPHCellList.jl, git status --short, git add src/SPHCellList.jl, and git commit -m "Inline half-step in neighbor loop".
• No automated test suite or runtests was executed as part of this change, so no automated tests were run or reported here.

---

Codex Task