fix(eval,cookbooks,ci): harden evaluation streaming, multi-node checks, and preview generation

[umran666]

While running through the evaluation benchmarks and fine-tuning cookbooks, I ran into a few edge cases and stability issues where transient network hiccups, gateway idiosyncrasies, or missing env vars caused jobs to fail or abort before writing aggregate results.

• UniGenBench Gateway Streaming & Teardown:
  • Skips empty choices chunks emitted by some OpenAI-compatible VLM gateway endpoints (e.g., terminal usage-only frames) to avoid raising IndexError mid-stream.
  • Adds exponential backoff between retries so rate-limited gateways aren't hammered.
  • Explicitly closes the underlying async HTTP client on vlm.close() to prevent lingering connection warnings upon interpreter exit.
• UniGenBench Judge Tuning & Metrics:
  • Exposes CLI args --vlm_timeout (default 300s, up from hard-coded 100s) and --vlm_max_retry (default 3) so slower judge completions aren't prematurely dropped and re-prompted.
  • ugb_scorer.py now defensively parses both native python lists and legacy stringified list representations.
  • Samples whose retries are exhausted are recorded under a "failed_samples" key in the result JSON (alongside success_count) instead of being silently omitted.
• PAIBench-G Retry & Partial Results:
  • Added per-video retry (VIDEO_RETRIES = 1) so transient AMT scoring failures get a second attempt before failing the run.
  • If a worker or video scoring fails, the script now writes whatever was cleanly scored and records the remaining entries under "missing_videos" in the result JSON, rather than aborting prior to the merge step.
• PAIBench-C GPU Clamping & Docs:
  • Parses CUDA_VISIBLE_DEVICES to clamp evaluation GPU workers to the actual number of visible devices, preventing oversubscription when COSMOS3_NUM_GPUS exceeds the physical count.
  • Fixed contradictory quick-start comments in run_paibench_c.sh and added an explicit 1-sample smoke-test example.
• SFT Launch Script Guards:
  • Added fail-fast parameter validation in cookbooks/cosmos3/generator/action/finetune/*.sh when NNODES, NODE_RANK, or MASTER_ADDR are partially set, surfacing a clear error before torchrun's rendezvous layer fails.
• Atomic Video Previews:
  • preview_helpers.py now encodes to a sibling .partial.mp4 first and swaps atomically with os.replace, preventing corrupt/incomplete preview files if an interactive cell is interrupted.
• CI Workflow Guard:
  • Added an empty-array guard in .github/workflows/validate-notebooks.yml so ruff exits cleanly when no notebooks match the glob.
• Self-Checks:
  • Added evaluation/cosmos3/generator/unigenbench/query_core_selfcheck.py and evaluation/cosmos3/generator/paibench_g/merge_result_selfcheck.py (stdlib-only assert tests, requiring no GPU or network) to verify streaming chunk parsing and merge aggregation edge cases.