[Feature Request] Serve every native-spin DPA model through dpa4spin #6133

Summary

Native-spin DPA4/DPA4C models are meant to run through pair_style dpa4spin / dpa4spin/kk, which provide magnetic_force cartesian for Newtonian spin dynamics (#6077). Several native-spin models and run setups cannot use these styles yet and fall back to pair_style deepspin, which offers only the precession convention. This issue tracks the missing pieces so that every native-spin model can run through dpa4spin.

Detailed Description

1. Multi-block DPA4 models frozen with the graph lower. A native-spin model with more than one message-passing block exports the graph (NeighborGraph) lower together with a with-comm artifact, which carries the per-block border_op exchange of ghost features for multi-rank runs. DeepSpin routes every archive with has_comm_artifact to DeepSpinPTExpt, and dpa4spin serves only NativeSpinPTExpt. Such a model therefore runs only with deepspin, on one rank or several, and cannot use magnetic_force cartesian. In current master this covers every uncompressed multi-block DPA4 native-spin model. Proposed direction: carry the halo exchange in the main program of the native-spin graph lower, as the compact canonical lower already does (communicator and border plan as inputs, deepmd::border_exchange inside the program, an empty plan on one rank). Every native-spin archive then routes to NativeSpinPTExpt, and no native-spin archive needs a with-comm artifact.
2. Uncompressed DPA4 on dpa4spin/kk. dpa4spin/kk requires the compact canonical lower, which only a compressed DPA4C exports today. A single-precision DPA4/DPA4.5 model frozen on a CUDA device could export the same compact lower and be evaluated by its own graph forward, so it runs device-resident on one GPU or several.
3. Model deviation. dpa4spin and dpa4spin/kk take a single model: no ensemble on the pair_style line, and no out_freq, out_file, atomic, relative or relative_v. deepspin writes the deviation of the force and the magnetic force; dpa4spin should do the same, and deepmd/kk lacks it too.
4. Frame and atomic parameters. dpa4spin passes no fparam/aparam (nor their *_from_compute/*_from_fix variants), so a native-spin model that requires them runs only with deepspin. The compact canonical lower carries no frame or atomic parameters either, so such a model would use the graph lower of item 1 on the host style.

I plan to open pull requests for items 1 to 3; item 4 is open.

Further Information, Files, and Links

• Routing: native_spin_backend_serves in source/api_cc/src/DeepPotPTExptPlugin.cc.
• With-comm decision: _needs_with_comm_artifact in deepmd/pt_expt/utils/serialization.py.
• Compact canonical spin lower with the halo exchange in its main program: forward_lower_canonical_graph_exportable of the native-spin model in deepmd/pt_expt/model/native_spin_model.py.
• Related: [Feature Request] Native Cartesian magnetic-force mode for Newtonian spin dynamics in LAMMPS #6077 (Cartesian magnetic force), feat(lammps): support Cartesian native-spin magnetic forces #6084.