fix(optimizer): route ETP-only experts through expert groups

[guapisolo]

Depends on: #82
Stacked on the parent PR's branch. Review / merge the parent first.

Summary

Route ETP-only expert parameters through expert reduction and LayerWise ownership.

Symptom & Reproduction

• Symptom: At parent head 3d107754c, TP2/ETP1/EP1 expert parameters carry allreduce=True, so DDP puts them in dense buffers and LayerWise uses the dense ownership plane.
• Reproduction: Run NVIDIA_PYTORCH_VERSION=25.06 torchrun --standalone --nproc-per-node=2 -m pytest -q tests/unit_tests/test_muon_optimizer.py::test_real_moe_ddp_layerwise_muon_expert_ownership_tp2_etp1_ep1; the parent behavior violates the expert-tag and expert-buffer assertions, while this commit passes them.

Root Cause

1. ColumnParallelLinear derived param.allreduce from EP > 1 only.
2. DistributedDataParallel selected dense buffers for those mislabeled parameters.
3. _get_param_groups propagated dense ownership into LayerWiseDistributedOptimizer.

Fix

Port the producer predicate is_expert and (EP > 1 or ETP != TP) from NVIDIA/Megatron-LM#5916 at head 6ccd20e2005b0698e908d42db771e2f22b0281e8 across native, Transformer Engine, grouped, legacy, and ModelOpt linear layers. The TE helper also restores numbered-parameter TP metadata and preserves SwiGLU partition stride; DDP remains unchanged because it already honors param.allreduce.

Verification

• pytest -q tests/unit_tests/transformer/moe/test_grouped_mlp.py -k 'expert_parameter_attributes' tests/unit_tests/post_training/test_modelopt_module_spec.py -k 'expert_parameter_attributes or modelopt_expert_linear': 9 passed.
• The 2-GPU native, TE CPU-init, and ETP-only DDP collective set: 3 passed on each rank; the optional legacy GroupedMLP runtime case skipped because grouped_gemm is unavailable.
• The 2-GPU TE stride=2 plus real MoE end-to-end set: 2 passed on each rank.
• The 8-GPU real MoE → DDP → Muon/Adam → LayerWise test: all eight ranks passed exact buffer, owner, norm, clipping, update, and EDP replica assertions.
• git diff --check, isort --check-only, and python -m compileall -q: passed for the changed files.

Review Focus

• _set_expert_parameter_attributes: numbered weight/bias matching, row-bias replication, and partition_stride preservation.
• use_expert_pgs: identical topology semantics across every expert parameter producer.
• test_real_moe_ddp_layerwise_muon_expert_ownership_tp2_etp1_ep1: deterministic gradients still exercise real MoE construction, DDP collectives, optimizer grouping, ownership, clipping, and updates.

[guapisolo]

@codex review