fix(rollout): separate transfer and policy versions

[quinnarnold]

Summary

Keep receiver-local weight-update call counts separate from the policy version stamped onto rollout output. Concrete rollout engines no longer overwrite published trainer provenance while receiving weight buckets or transfers.

Current trainer and rollout-manager publication ordering from #310 and #315 remains unchanged.

Related Issue

Fixes #324.

Test Plan

• SKIP=no-commit-to-branch uvx pre-commit run --files unirl/rollout/engine/fastvideo/engine.py unirl/rollout/engine/sglang/engine.py unirl/rollout/engine/sglang_diffusion/engine.py unirl/rollout/engine/vllm_omni/engine.py --show-diff-on-failure — all applicable hooks passed.
• Disposable CPU/AST harness — verified transfer calls only advance the receiver-local counter, successful full publication assigns the trainer version afterward, failed publication does not advance provenance, and agentic forwarding stamps the published version.
• Not run: Ray/GPU model smoke test; Ray and compatible model hardware are unavailable locally.

Compatibility / Risk

No configuration, checkpoint, data-format, or public API changes. Batch and agentic publication contracts are unchanged; only the conflicting internal receiver counter is separated.

Reviewer Notes

Later upstream work already implemented the trainer-to-worker version publication requested in #324, so this PR contains only the remaining four-engine counter separation. Duplicate searches found no overlapping open PR; #390 is unrelated. AI assistance was used; the complete diff and validation results were reviewed before publication.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.