[Bugfix][Router] Keep avg_latency within the request-stats sliding window

[David-Wu1119]

RequestStatsMonitor.get_request_stats reports per-engine statistics over the --request-stats-window sliding window. Before reading the QPS and TTFT monitors, it calls update_no_value(current_time) on each one so that samples older than the window are dropped. It never did that for the latency monitor, so avg_latency only shed old samples when a new request finished on that engine. Once an engine went idle, avg_latency stayed at the last value forever: the router kept exporting it (vllm:avg_latency) and logging it, while QPS and TTFT for the same engine had already dropped to 0 / -1.

For example, with a 10 s window, take one request that finishes 2 s after it started. 100 s later the stats read qps=0.0, ttft=-1, avg_latency=2.0. This change makes it avg_latency=-1, the same as TTFT, by dropping old latency samples the same way. (Open #1054 adds the same cleanup for avg_decoding_length; this is the latency counterpart.)

Tests:

• New test_avg_latency_drops_requests_outside_the_window in src/tests/test_request_stats.py checks the example above. It fails on main (avg_latency is still 2.0) and passes with this change.
• pytest src/tests: 241 passed, Python 3.12, pip install -e . plus pytest, pytest-asyncio and httpx.
• pre-commit run on both files passes.

Found while auditing the router's request statistics with AI assistance; the fix and test were written with Claude Code.

---

• Make sure the code changes pass the pre-commit checks.
• Sign-off your commit by using -s when doing git commit
• Try to classify PRs for easy understanding of the type of changes, such as [Bugfix], [Feat], and [CI].

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[gemini-code-assist]

Code Review

This pull request ensures that requests outside the sliding window are correctly dropped when calculating average latency. This is achieved by calling update_no_value on the latency monitors within get_request_stats. Additionally, a new unit test test_avg_latency_drops_requests_outside_the_window has been added to verify this behavior. No review comments were provided, and the changes appear correct and complete.