feat(rollout): support deterministic concurrent rollouts

[Le8r0nJames]

Description

This PR makes OpenAI-proxy rollouts reproducible without serializing grouped
samples or limiting SGLang to one running request.

It:

• derives stable per-request sampling seeds from the logical session identity
  and request index, while preserving explicitly supplied seeds;
• carries sample_idx through grouped workflow contexts and merges concurrent
  samples in canonical order;
• adds seed support to the Chat Completions and Responses APIs and forwards it
  to SGLang as sampling_seed;
• exposes SGLang deterministic inference configuration and warns when the
  selected attention backend is not documented as supported;
• orders completed rollout tasks by task ID without random shuffling in
  deterministic mode;
• warns when deterministic sampling is combined with
  max_head_offpolicyness > 0, because end-to-end deterministic training
  requires synchronous rollout generation.

The feature remains opt-in. Existing rollout behavior is unchanged when
deterministic_sampling is disabled.

Related Issue

N/A — no public issue is currently linked.

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 💥 Breaking change
• 📝 Documentation update
• ♻️ Refactoring
• ⚡ Performance improvement
• ✅ Test coverage improvement

Checklist

• I have read the
  Contributing Guide
• Pre-commit hooks pass (pre-commit run --all-files)
• Relevant tests pass; new tests added for new functionality
• Documentation updated (if applicable; built with ./docs/build_all.sh)
• Branch is up to date with main
• Self-reviewed via /review-pr command
• This PR was created by a coding agent via /create-pr
• This PR is a breaking change

Breaking Change Details (if applicable):

N/A.

Additional Context

SGLang deterministic inference provides batch-invariant execution for the
documented flashinfer, fa3, and triton attention backends. AReaL still
needs stable request seeds and canonical host-side ordering so concurrent
grouped samples remain reproducible.

For end-to-end deterministic training, users should configure:

rollout:
  deterministic_sampling: true
  max_head_offpolicyness: 0

sglang:
  enable_deterministic_inference: true

max_head_offpolicyness=0 does not serialize requests within a rollout batch.
Grouped samples and prompt workflows remain concurrent; it only disables
cross-step rollout/training staleness.

Validation performed:

• 20 focused deterministic-sampling tests passed in the development image;
• Ruff, formatting, SPDX, generated CLI documentation, and diff checks passed;
• a 10-step Qwen3-0.6B colocated A/A run with batch size 16 and 8 samples per
  prompt produced identical 1,280 rollout records and bitwise-identical
  non-timing training scalars;
• a frontier-only ablation matched both A/A arms and the deterministic
  reference exactly, confirming that deterministic task ordering and disabling
  shuffle are sufficient for the validated synchronous, no-rejection path.

Automatic seed derivation currently applies to OpenAI-proxy sessions. Callers
that issue semantically concurrent sibling requests within one session should
provide explicit stable seeds if they require a fixed request-to-seed mapping.

[sitabulaixizawaluduo]

Thank you for your contribution. A similar submission has already been made. Please go to another PR for review and submit your comments.