[Review only][CPU] Assemble auxiliary outputs on the execution path

[aoshen02]

Review-only alternative: CPU execution-side assembly

Alternative to #13, not an additional upstream fix. Both target r3_offload;
do not merge the alternatives together. Duplicate-work checks found upstream
vllm-project#45635 and the existing fork #13; this alternative deliberately offers
a different lifecycle tradeoff for review, not a competing upstream PR.

Finish R3 processing inside sample_tokens() before returning. The next worker
RPC cannot clean request state until this completes. Keep the existing async
GPU-to-CPU copy, wait for it after draft proposal, then assemble on CPU. Remove
the pending-output event/wait handshake; output consumption no longer mutates
the connector. No new output type or no-forward wrapper.

Tradeoff: CPU assembly and its copy wait enter the execution critical path.
Background storage remains unchanged.

The runner calls connector-local worker.finish_aux_output(); completion
implementation is no longer added to the public AsyncOutput class. Replace the
retained step metadata object with a presence flag and reuse the computed end
position.

Validation

• /home/aoshen/vllm/.venv/bin/python -m pytest tests/distributed/aux_output_connector -q: 58 passed (16.83s). Includes completion-order coverage.
• Full pre-commit hooks on changed files passed during commit.
• Independent Claude Code review agrees with the execution-order proof.
• GB200 UniProc with VLLM_GPU_SYNC_CHECK=error: 73 store/config tests and both HTTP R3 tests pass (Slurm 29066).
• Raw-snapshot-to-HTTP oracle on 0ed49a1: 64 requests, 16,778,176 prompt/decode R3 rows match exactly (Slurm 29059). This does not independently validate model routing computation or D2H.
• Real MTP: 32 requests complete; 2,715 draft tokens, 93 accepted, 2,622 rejected. R3 shape/range checks pass (Slurm 29061), not an independent MTP value oracle.
• Forced memory-pressure preemption: 64 requests complete with 167 preemptions and R3 shape/range checks passing (Slurm 29080 CPU stage).
• Same-node GB200 throughput on 0ed49a1, 256K prompt / concurrency64 / one output token: 82,375 and 82,312 tokens/s, mean82,343, 2.73% below native image R3 (84,652). Actual weights: DeepSeek-V4-Flash-Base; TP1/DP8/EP. This is primarily prefill throughput. Both chunked/short prefix-reuse checks pass; newly computed cold/hot decode rows are not claimed numerically identical.
• CPU/GPU finish-path profiles completed (Slurm 29093). They identify GPU per-block publication copies as a batching target, but do not isolate CPU's lost async overlap from storage/assembly overhead. Summed worker/event-wait times are not end-to-end regression percentages.

AI assistance (Codex and Claude Code) was used. For human line-by-line
review and hardware validation; not ready to merge.