Add PhysicsNeMo domain-parallel mesh sharding support

[rmclaren]

Motivation

• Enable domain (mesh) parallelism so the global icosahedral mesh can be split across GPUs using PhysicsNeMo ShardTensor primitives, reducing per-GPU memory and enabling larger models and datasets.
• Integrate mesh-parallel sharding with the existing data-parallel bin sampling and multi-node DDP training so existing sampling/resampling flows continue to work.

Description

• Add a new utility module gnn_model/domain_parallel.py that provides MeshShard, build_mesh_shard, filter_mesh_edges, filter_bipartite_edges, shard_tensor_dim0, init_domain_parallel_context, and maybe_build_shardtensor to manage simple mesh splits and optional PhysicsNeMo ShardTensor construction.
• Wire domain-parallel options into the data module by adding domain_parallel flags to GNNDataModule and applying mesh sharding in setup() and graph construction (_create_graph_structure) with filter_mesh_edges and filter_bipartite_edges.
• Add distributed-aware sampling by using torch.utils.data.distributed.DistributedSampler in train_dataloader() and val_dataloader() so bins are split cleanly across ranks when DDP is active.
• Wire domain-parallel CLI/config options into train_gnn.py (--domain_parallel, --domain_parallel_mesh_shape, --domain_parallel_mesh_dim_names) and pass them into GNNLightning and GNNDataModule, and add an on_fit_start hook in GNNLightning to build the mesh shard and optionally convert sharded local mesh tensors into PhysicsNeMo ShardTensors.
• Document usage and installation notes in gnn_model/README.md, including the new flags and optional pip install physicsnemo step.

Testing

• No automated tests were executed as part of this change and no CI/test run was performed.

---

Codex Task