Use per-particle interaction accumulators

[AhmedSalih3d]

Motivation

• Reduce synchronization and per-thread reduction overhead by accumulating interaction contributions locally per particle instead of writing into threaded reduction buffers.
• Avoid repeated zeroing/reset of large reduction arrays between neighbor loops by writing each particle once after its local accumulation.
• Keep kernel-output and shifting behavior optional and zero-cost when disabled via SimulationMetaData dispatch.
• Improve cache locality and simplify the simulation loop by moving accumulation into the per-particle traversal.

Description

• Added construct_full_stencil and a KernelOutputLocal! pair to support per-particle local accumulation of kernel outputs when requested.
• Implemented NeighborLoopPerParticle! overloads (with and without shifting) that iterate per particle, locally accumulate dρdt, acceleration, kernel values, and shifting terms, and write them once per particle.
• Implemented ComputeInteractionsPerParticle! variants that compute pair contributions and return local increments instead of mutating threaded reduction arrays.
• Switched SimulationLoop / RunSimulation to use the new per-particle neighbor loop and removed the intermediate threaded reduction/reset calls and the SimThreadedArrays reduction steps for the main interaction pass.

Testing

• Ran julia --project=. -e 'using Pkg; Pkg.test()', which was attempted but precompilation stalled during CSV/UnicodePlots precompilation and the run was interrupted.
• Per AGENTS.md, the repository currently lacks a test directory so Pkg.test() will report an error until tests are added, and full automated tests were therefore not completed.
• No additional automated test suite was available to validate numerical equivalence after the change, so manual/bench testing is recommended before production use.
• Command(s) executed: julia --project=. -e 'using Pkg; Pkg.test()' (precompilation stalled/interrupted).

---

Codex Task