… pt) (deepmodeling#5555)

## What

Completes the DPA4/SeZM **SO3 grid projection** port to the dpmodel
backend so it faithfully **mirrors master's current pt**
`sezm_nn/grid_net.py`. After this, the flagship
`examples/water/dpa4/input.json` (which sets `ffn_so3_grid=true`,
`message_node_so3=true`, `grid_mlp`) runs on dpmodel/pt_expt.

Builds **on top of** the S2-grid base that deepmodeling#5517/deepmodeling#5552 landed
(`GridProduct`/`GridMLP`/`op_type='mlp'`, the `_project_frames`
refactor). Master's dpmodel was the S2 (`n_frames==1`) slice with
SO3/cross-mode fail-fast guarded; this PR generalizes those ops to
frame-aware (`n_frames>1`) + cross-mode and adds the missing SO3 pieces
— matching current pt exactly (single source of truth: dpmodel == pt).

Supersedes deepmodeling#5547 (which ported the *pre*-deepmodeling#5552 design and went
structurally stale).

## Changes (all mirror current pt)

- **`grid_net.py`**: add `_project_frames`; generalize
`GridMLP`/`GridBranch` to frame-aware (`n_frames`); generalize
`BaseGridNet` (un-guard `mode='cross'`, `layout='flat'`,
`residual_scale_init`, `n_frames>1`; frame-axis to/from-grid via
`xp.matmul`+reshape); add
`FrameContract`/`FrameExpand`/`_build_frame_degree_index`; add
`SO3GridNet` (self+cross).
- **`projection.py`**: add `SO3GridProjector` (Wigner-D quadrature) +
`resolve_so3_grid`/`_build_so3_frame_set`.
- **`ffn.py`**: un-guard `ffn_so3_grid` → `SO3GridNet(mode='self')`.
- **`so2.py`**: un-guard `node_wise_{s2,so3}`/`message_node_{s2,so3}` →
cross-mode grid products, applied in `call` + round-tripped in
serialize.

## Validation

- Component parity vs pt (weight-copied fp64): `_project_frames`,
`GridMLP`/`GridBranch` (incl. S2 byte-identical regression),
`BaseGridNet` cross/flat/residual, `FrameContract`/`FrameExpand`,
`SO3GridProjector` matrices, `SO3GridNet` self+cross (op_type
glu/mlp/branch, kmax 1&2) — all **1e-12**; rotation equivariance
**1e-10**.
- **fp32** grid-path parity at the computation-in-fp32 budget (actual
diffs 1e-6–1e-8 ≪ 1e-4).
- Full-descriptor pt→dpmodel via
`DescrptDPA4.deserialize(pt.serialize())` on the example config (lmax=3,
mmax=1) — **~1e-14** — proving `dp convert-backend` schema interop.
- Permutation-invariance + masked-edge no-op.
- Cross-backend consistency rows (pt vs dpmodel **and pt_expt**,
mixed_types) for ffn_so3_grid / message_node_so3 / both / grid_mlp.
- **Verified on remote GPU (Tesla T4):** 617 (grid+parity+pt_expt) + 50
(consistency) pass, no CUDA device errors.

pt_expt forward works today via auto-wrap (consistency + descriptor trio
green) — no explicit registration needed.

## Known limitations

- pt_expt **training** Parameter-promotion for the new weight-bearing
grid classes, `torch.export`/AOTI grid coverage, training e2e, argcheck
`doc_only_pt_supported` removal, and freeze/DeepEval are a **follow-up
PR**.
- `grid_method='e3nn'` (non-Lebedev product grid) stays fail-fast
(Lebedev-only, per parent design).
- fp32 grid paths use a ~1e-4 budget by design; fp64 is the parity
reference.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Added SO(3) grid projection support and frame-aware grid networks for
DPA4 descriptors, including SO(3)-based FFN and improved cross-mode
grid-product wiring.
* Extended grid modules to support multi-frame configurations with
per-degree frame mixing, and added full SO(3) projector/network
serialization.
* **Bug Fixes**
* Enabled previously disabled/unsupported DPA4 SO(2) convolution
cross-mode SO(3)/S2 grid products.
* **Documentation**
* Updated DPA4 porting-layer documentation to clarify supported
configuration flags.
* **Tests**
* Added/expanded parity, equivariance, serialization/roundtrip, and
torch-namespace compatibility tests for the new SO(3) and frame-aware
paths.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Han Wang <wang_han@iapcm.ac.cn>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>