Motivation

• Avoid cross-thread writes and reduce overhead when kernel output and shifting are disabled by iterating per particle with per-thread accumulators.
• Provide a lighter-weight interaction path that skips kernel recording and shifting bookkeeping for the common NoShifting/NoKernelOutput configuration.

Description

• Add an overload of NeighborLoop! specialized for SimulationMetaData{..., NoShifting, NoKernelOutput, ...} that iterates per particle, partitions work into stable per-thread chunks, and accumulates contributions into per-thread buffers (SimThreadedArrays.dρdtIThreaded and AccelerationThreaded).
• Implement ComputeInteractionsPerParticleNoKernel! which computes density-change and acceleration contributions between a particle pair but omits kernel-output and shifting side-effects, and returns the updated accumulators.
• Use SimParticles.Cells, ParticleRanges, and CellDict to lookup same-cell and neighbor-cell particle ranges and perform the full stencil neighbor traversal while skipping the center cell when iterating neighbors.
• Changes are contained in src/SPHCellList.jl (new specialized NeighborLoop! and ComputeInteractionsPerParticleNoKernel!).

Testing

• Ran the package test command: julia --project=. -e 'using Pkg; Pkg.test()', but the run did not complete because package precompilation (UnicodePlots) stalled and the process was interrupted, so the test suite did not finish successfully.
• No other automated tests were executed in this rollout.

Codex Task