feat(novita): 1h async media poll budget for a single generation

[duanbing]

What

Splits the overloaded REQUEST_TIMEOUT constant in the Novita provider:

• REQUEST_TIMEOUT (300s) — stays the per-HTTP-request timeout (create-task submit, single poll fetch).
• ASYNC_TASK_TIMEOUT (3600s / 1h) — new constant, the total poll-loop deadline for one async media generation.

Why

Heavy models (e.g. sora_2_pro_i2v) routinely render past the old 300s budget, producing spurious Novita async task … did not complete within 300s failures even though the upstream render eventually succeeds.

On the RouterBase side this generation is now single-attempt, no-retry (video_submitter MAX_ATTEMPTS=1), so this is the one and only budget for a request — one upstream Novita task, no throwaway re-submissions spawning fresh task ids.

Paired change

Requires the RouterBase parent PR that:

• pins this submodule commit,
• sets MAX_ATTEMPTS=1 and raises the worker reaper above 1h (STUCK_TIMEOUT_MINUTES=70),
• adds 60s TCP keepalive on the backend→gateway client (the connection sits idle for up to 1h while the gateway polls Novita),
• surfaces in-flight generations in /logs.

Risk

A single submission now holds an idle connection to the gateway open for up to 1h. Backend-side keepalive mitigates conntrack/idle drops; worth watching connection counts at high concurrency on first real 1h render.

🤖 Generated with Claude Code

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the Contributor License Agreement (CLA) and hereby sign the CLA.

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}