[codex] Add Empire AI NVL72 training launcher

[jder]

Summary

• add an Empire AI Beta NVL72 Slurm/Pyxis launcher for half-degree OM4 training
• use the public m2lines ARM64 container manifest digest that works with Empire AI's Pyxis/Enroot stack
• default W&B runs to the ocean_emulators entity instead of the stale samudra entity

Context

The NVL72 nodes use ARM CPUs and the cluster's Pyxis/Enroot stack did not import GHCR OCI index tags reliably, so the launcher pins the ARM64 manifest digest. The script supports a 2-GPU smoke mode and a 9-node / 36-GPU full mode, validates the expected data layout under ~/data/om4_halfdeg, and passes W&B entity/project overrides explicitly.

Validation

• bash -n scripts/empireai_nvl72_train.sbatch
• uv run python - <<'PY' ... WandBConfig().entity == "ocean_emulators" ... PY
• pre-commit during commit: ruff, ruff-format, mypy, validate-schemas, detect-secrets, reuse lint
• Empire AI smoke job 20654: 2 GPUs on 1 node, completed 0:0
• Empire AI smoke job 20673: 8 GPUs on 2 nodes, completed 0:0