fix(rollout): preserve sampled EOS probabilities

[bvolpato]

Single-assistant-message rollouts remove even a sampled final EOS, then append it with logprob 0.0 and an active loss mask. An unchanged policy with EOS probability 0.01 consequently gets a sequence importance ratio of 0.01 instead of 1; outlier filtering can reject the whole on-policy trajectory.

Keep the sampled EOS and its probability on the final environment turn. Intermediate EOS tokens are still removed while the assistant message continues. If finalization adds a genuinely synthetic EOS, exclude it from policy/importance-correction loss and leave the reward on the last generated token. Empty and EOS-only responses are covered as well.

Reproduction and validation

uv run --isolated --extra skyrl-train --extra dev pytest \
  tests/train/generators/test_singleturn_eos_logprobs.py \
  tests/train/generators/test_skyrl_gym_generator.py -q

• 40 tests passed in the isolated Python 3.12/PyTorch CPU environment.
• The nine new regression cases produce 6 failures and 3 passing controls against the original generator, then all pass with the fix.
• The actual agent loop and importance-correction functions verify ratio 1 and acceptance for an unchanged policy, including an unlikely EOS. Cases cover intermediate turns, route metadata, synthetic EOS, missing logprob capture, empty/EOS-only output, and context/length truncation.
• An independent check through production tensor preprocessing, advantage normalization, and PPO/TIS gives sampled EOS gradient -0.5 and synthetic EOS gradient 0, with total reward preserved.
• Ruff, Black, and gitleaks hooks passed.
• GitHub GPU CI stopped during Anyscale setup because the fork job has no ANYSCALE_CLI_TOKEN; it did not execute tests. Upstream code-quality and gym checks passed; remaining CPU jobs were still running when submitted.

The existing executor timing test stalled in the sandbox but passed outside it, along with the complete 40-test suite.

The sample-support work in #2082 preserves the EOS support row but retains the zero rollout logprob; this fix addresses that separate bookkeeping defect.

Downsides

Synthetic EOS tokens no longer receive a policy gradient or terminal reward. This corrects the objective but can change loss and learning curves for existing single-assistant-message runs that previously trained on those fabricated actions.

Risk and rollback

The affected path is non-batched generation with use_conversation_multi_turn=false. Reward placement and token alignment are the main regression risks; tests cover final observations, intermediate turns, truncation, and routed-expert metadata. Revert this commit to roll back the behavior.