fix(pd-lm): SIGPIPE-proof the driver-version detection in the node-setup script

[claude-spd1]

Description

The generated sbatch node-setup script detects the CUDA extra via
nvidia-smi --query-gpu=driver_version --format=csv,noheader | head -n 1 | cut -d. -f1.
nvidia-smi writes one line per GPU; head -n 1 closes the pipe after the first line, and under the script's set -o pipefail a late write from nvidia-smi becomes SIGPIPE → exit 141, killing the whole srun step before the workspace even builds. It's a per-node race — observed three times across 1- and 4-node submissions (jobs 1050555, 1050877's requeue, 1090526), and with N nodes per job the odds multiply.

Fix: query GPU 0 alone (nvidia-smi --id=0 ...) — cut consumes the full stream, no mid-stream pipe close, same DRIVER_MAJOR.

Motivation and Context

Flaky launch failures burn queue time (an opportunistic 4-node allocation dies in 5s and goes back to the end of the line) and look like real breakage.

How Has This Been Tested?

Patched the submitting checkout and resubmitted the same 4-node probes; generated scripts carry the new line. nvidia-smi --id=0 --query-gpu=driver_version --format=csv,noheader verified to emit exactly one line on an 8-GPU node.

Does this PR introduce a breaking change?

No — submit-side script generation only.

🤖 Generated with Claude Code

Crew-Address: task/scan-read-activations-taps

[claude-spd1]

Duplicate of #984 (same one-line --id=0 fix, opened earlier and verified with a 300-iteration repro: 57/300 failures unpatched, 0/300 patched — see task workspace-fetch-sigpipe-flake and lore:2026-07-15--finding--pd-lm-setup-sigpipe-nvidia-smi-race). Closing in favor of #984. For what it's worth this PR's three independent hits (jobs 1050555, 1050877's requeue, 1090526 — 1- and 4-node submissions on cw-rno2) are further field confirmation of the diagnosis.