[fsspec-aware #479] cgns: remote-aware init_from_disk (read, priority 4)

[xroynard]

Sub-task of #479. Read side, cgns backend — hardest read task. Covers part ④ of the read-side analysis (issuecomment-5166383239).

Goal

Allow cgns/reader.py::init_datasetdict_from_disk to read from a remote target.

Current state (verified on main)

• init_datasetdict_from_disk builds CGNSDataset(path).
• CGNSDataset.__getitem__ → Sample.load_from_dir → Sample.load → CGM.load(str(meshes_dir / "mesh_*.cgns")).
• pyCGNS (CGNS.MAP.load) reads the CGNS/HDF5 file through a native C file handle — no fsspec object accepted.
• Split discovery also uses local_path.iterdir() (same local-only issue as the zarr reader).
• No streaming reader exists for cgns today.

Options (design choice, not a config flag)

• (a) Materialize-to-temp: download remote sample dirs into a local temp dir, then reuse the existing local reader unchanged. Simpler, bounded temp usage (~1 sample at a time if done lazily in __getitem__).
• (b) fsspec rewrite: rework Sample.load around fsspec. More invasive, touches containers/sample.py.

Recommend evaluating (a) first (lazy per-sample temp in __getitem__) as it isolates the change to the cgns reader.

Depends on

• [fsspec-aware #479] common: remote-aware load_constants_from_disk (read, priority 3) #487 (remote constants) for a complete remote read.
• Split discovery fix (fsspec fs.ls) shared with the zarr reader approach in [fsspec-aware #479] zarr: make init_from_disk read path fsspec-aware (read, priority 2) #486.

Tests

• memory:// or local-temp-backed source: verify a sample loads correctly.

Priority

★★★ (4/5) — most invasive read task; do after zarr read (#486) and constants (#487).