feat(dp): add GPU-resident edge compute path for DPA4/SeZM models

[SchrodingersCattt]

Summary

Add a fully GPU-resident edge-based compute path (compute_gpu_edges) for DPA4/SeZM models in the Deep Potential integration, eliminating all host-device data transfers per MD step for large-box simulations.

Modification

1. Replace DP_Data cell-list with Neighbor class — uses skin-based amortized neighbor-list rebuild, consistent with other potentials in GPUMD.

2. Add compute_gpu_edges() device path — when cutoff < L/2 for all periodic directions:

   • Builds global neighbor list at rc + skin on GPU (amortized)
   • Filters to local list within rc per step
   • Materializes compact [2, nedge] edge schema via dp_fill_edges kernel
   • Calls deep_pot.compute_edges_gpu() (new DeePMD-kit device API)
   • Scatters outputs via dp_scatter_outputs kernel
   • Zero host-device copies per step

3. Simplify fallback path (small boxes where cutoff >= L/2):

   • Delegates PBC handling entirely to DeePMD (deep_pot.compute() without nlist)
   • Removes explicit ghost atom construction (caused force inconsistency for message-passing networks)
   • Handles non-periodic dirs by inflating box (orthogonal + triclinic)

4. Remove profiling infrastructure (DP_Profile_Record, env-var GPUMD_DP_PROFILE) — re-addable in follow-up.

Dependencies

• Requires DeePMD-kit with compute_edges_gpu() C++ API (feat(dpa4): multiple updates for dpa4 deepmodeling/deepmd-kit#5734, merged 2026-07-07).

Others

Co-authored-by: OutisLi (DeePMD-kit device API author)