SkyRL recipes: 4-node / 32B speedup numbers not reproduced locally

[sfc-gh-kganesan]

Summary

The 4-node / 32B launchers and their speedup claims have not been
independently reproduced in this validation — the validation box is a single
node (1×8 H200), so the multi-node paths were exercised only by inspection.

Unverified locally

• txt2sql/run_qwen3_32b_bird_grpo_arl_4node.sh +
  run_qwen3_32b_bird_grpo_fsdp_4node.sh
• long_context_qa/run_qwen3_32b_loongrl_grpo_arl_4node.sh +
  run_qwen3_32b_loongrl_grpo_fsdp_4node.sh
• The ~2.17× / ~2.38× 32B wall-clock numbers quoted in the recipe READMEs
  are the PR author's measurements, not reproduced here.

What was validated (single node, 8B)

• LoongRL 8B ARL vs FSDP A/B: 1.39× e2e (1.55× train, 2.18× generation).

Acceptance

• Reproduce a 4-node 32B ARL-vs-FSDP A/B on a multi-node cluster.
• Confirm (or correct) the 32B speedup numbers in the READMEs against a
  fresh run.