perf(router): stream the data path instead of buffering it

[Rockdu]

What

Streams the router's data path. Three commits: submit the parser task off the event loop, strip hop-by-hop headers, and relay response bodies chunk-through instead of buffering.

Draft: stacked on feat/rollout-request-timing (the instrumentation these commits modify); rebase and mark ready once that branch lands.

Why

The router buffered every response twice (httpx aread() + the transport's bytearray) and serialised the two hops — nothing reached the client until the worker finished sending. At H3's 1.4 GiB responses that made router egress the largest cost in an admitted request (47.7%, above the denoise's 24.5%) and RSS peaked at 51 GiB.

Validation

Isolated harness (stub upstream -> this router -> N clients, 1384 MiB payloads), buffered vs streamed:

N	aggregate MiB/s	ttfb	router RSS
1	249 -> 501	2.7 s -> 0.00 s	4.2 GiB -> 57 MiB
16	340 -> 416	27.2 s -> 5.8 s	34.4 GiB -> 144 MiB

In the 17-GPU H3 run (with the sibling PRs): per-request egress median 61.6 s -> 4.5 s, rollout-phase GPU 35% -> 82%, train_wait 1010-1140 s -> 486-499 s.

Measured constraint — do not merge for raw float payloads alone. An ablation on the same run with --rollout-video-dtype keep (1384 MiB bodies, streaming on) REGRESSED: egress median 95.1 s vs 61.6 s buffered, train_wait 1260 s vs the 1010-1140 s buffered baseline. Buffering acted as a large elastic buffer between the engine-side and client-side rates; streaming replaces that with end-to-end min-rate flow, and at ~90% utilisation of the path the queueing dominates. The harness's +22%..2x was measured with an idle client loop and does not transfer. Streaming's wins at H3 payloads are memory (router RSS 34 GiB -> 144 MiB) and headroom; the throughput win requires the payload cut (#198's uint8) to land with it.

Known issue surfaced by the harness, pre-existing and not introduced here: uvicorn's 5 s keep-alive close races a client reusing a pooled connection (probability grows with N; disappears with keep-alive off). Worth a follow-up on its own.

Checklist

• pre-commit run --all-files passes — run on touched files
• Added/updated tests for new behaviour — no new tests; validated by the harness above
• pytest -x is green — tests/fast on the full stack, 215 passed
• If launch flags changed, --help still parses — no flag changes
• CLI reference docs — n/a
• Example walkthrough — n/a