feat: add vLLM GB200 GSM8K repro configs

[alec-flowers]

Summary

Adds a reproducible vLLM DeepSeek-V4-Pro GB200 GSM8K eval path and the two exact 1P1D configs used for the SA smoke comparison.

• Adds a self-contained benchmark.type: lm-eval runner that evaluates OpenAI-compatible chat endpoints with EleutherAI lm-eval.
• Runs the eval in the same runtime container as the server, matching the InferenceX/srt-slurm multi-node flow.
• Vendors the GSM8K task YAML, score threshold, and score validator used by the repro configs.
• Adds two exact vLLM GB200 1P1D eval recipes:
  • DEP8 prefill + TP8 decode: disagg-gb200-1p1d-dep8-tp8-gsm8k-smoke.yaml
  • DEP8 prefill + DEP8 decode: disagg-gb200-1p1d-dep8-dep8-gsm8k-smoke.yaml

SA Validation

Successful runs on the SA GB200 cluster:

• TP8 decode repro: job 15558
  • exact_match,strict-match: 0.8635
  • exact_match,flexible-extract: 0.8597
• DEP8 decode comparison: job 15559
  • exact_match,strict-match: 0.9636087945413192
  • exact_match,flexible-extract: 0.9628506444275967

Both use full GSM8K test split (1319 examples), 5 shots from the task YAML, EVAL_CONC=128, max_length=9472, and max_tokens=5376.

Test Plan

• PYTHONPATH=src /home/aflowers/Documents/agent/srt-slurm-gsm8k-worktree/.venv/bin/python -m pytest tests/ -q
  • 639 passed, 2 skipped, 6 deselected
• /home/aflowers/Documents/agent/srt-slurm-gsm8k-worktree/.venv/bin/ruff check src/srtctl tests/test_benchmarks.py
• Parsed both new recipes with load_config; both resolve as benchmark=lm-eval, served model deepseek-ai/DeepSeek-V4-Pro, EVAL_CONC=128.
• Ran the repo copyright-check logic locally; all checked files have NVIDIA SPDX headers.

Note: a fresh uv run pytest ... in the new worktree could not resolve jinja2 from the configured package index, so validation used the existing populated srt-slurm dev venv with PYTHONPATH=src pinned to this worktree.

Preflight Note

This PR also keeps srtctl apply from running blocking preflight implicitly. Some GB200 clusters keep model paths on per-compute-node local storage, so login-node preflight can fail even though the job would run correctly once scheduled. Explicit validation remains available via srtctl preflight -f ....