Add LeRobot pi0-FAST DROID multi-node training test case

[KeitaW]

Summary

• Adds a new test case at 3.test_cases/pytorch/lerobot/pi0-fast-droid/ for multi-node distributed training of a Vision-Language-Action (VLA) policy
• Uses LeRobot v0.4.3 with the pi0-FAST architecture (SigLIP vision + Gemma 2B language backbone) on the DROID 1.0.1 dataset (76k+ real-robot trajectories, 1.7TB)
• Multi-node launch via HuggingFace Accelerate over Slurm + Enroot/Pyxis with EFA networking

Files

File	Purpose
Dockerfile	NGC PyTorch 25.08 + EFA 1.47.0 + NCCL 2.28.4 from source + aws-ofi-nccl 1.14.1 + LeRobot[pi]
Makefile	Build → clean → enroot import
README.md	Full docs with architecture diagram, config tables, scaling variants, troubleshooting
slurm/run.sh	Multi-node sbatch script (Accelerate + Pyxis, HyperPod auto-resume)
requirements.txt	Accelerate + wandb
env_vars.example	Template for HF_TOKEN, output dir, node count
test_pi0_fast_droid.py	Pytest: docker build, smoke test, CLI entrypoint, EFA check
src/__init__.py	Package placeholder

Version compliance

• EFA 1.47.0 >= 1.47.0 ✓
• NCCL 2.28.4 >= 2.28 ✓
• CUDA 13.0 (via NGC 25.08-py3) >= 13.0 ✓

Test plan

• Verify Docker image builds successfully on a build node
• Run single-node training smoke test (1 node, small batch, few steps)
• Run multi-node training (2+ nodes) and verify all ranks join correctly
• Verify EFA is used for NCCL communication (NCCL_DEBUG=INFO output)
• Verify DROID dataset streaming/caching works from shared filesystem
• Run /ad-helper:lint-contribution — all checks pass
• Run /ad-helper:check-versions — all versions meet CI minimums