fix(deepswe): Configure default EOS_TOKENS and plumb into rollout launchers

[andytwigg]

Summary

This PR configures default EOS_TOKENS and plumbs them to rollout nodes for DeepSWE distributed RL runs (both local and GKE k8s), aligning with the pattern established in math_gsm8k_dist.

Problem

• In tunix/experimental/examples/deepswe_dist/launcher.sh, EOS_TOKENS defaulted to empty string (${EOS_TOKENS-}), so --eos_tokens was never passed to run_rollout_node.main unless manually set in the environment.
• In tunix/experimental/examples/deepswe_dist/k8s_launcher.sh, EOS_TOKENS was neither declared nor passed in start_rollout().
• Qwen models (Qwen3-1.7B, Qwen3-4B, Qwen3.5-35B-A3B) finish assistant turns with <|im_end|> (token 151645), while <|endoftext|> is token 151643. Without 151645 configured as an EOS token, vLLM fails to stop at the end of turns during rollouts, generating up to MAX_RESPONSE_LENGTH tokens (e.g. 61,440 tokens), hallucinating additional turns, and causing trajectory timeouts.

Changes

1. tunix/experimental/examples/deepswe_dist/launcher.sh:
   • Set default EOS_TOKENS=${EOS_TOKENS-'151645,151643'}.
   • Display eos tokens in startup banner.
2. tunix/experimental/examples/deepswe_dist/k8s_launcher.sh:
   • Define and export default export EOS_TOKENS=${EOS_TOKENS-'151645,151643'}.
   • Plumb ${EOS_TOKENS:+--eos_tokens=\"${EOS_TOKENS}\"} \ into run_rollout_node.main startup command.

Verification

• Validated bash syntax (bash -n).
• Verified generated rollout JobSet manifests via --dry-run:
  • Defaults to --eos_tokens="151645,151643".
  • Setting EOS_TOKENS="" omits --eos_tokens.
  • Setting custom EOS_TOKENS is correctly passed.