Add request latency histogram metric

[mikmatko]

Solves #871

Note: AI implemented. I'm not familiar with Rust. I've reviewed the code. I apologize if this is seen as too much AI slop :)

Expose granian_request_duration_seconds, a per-worker Prometheus histogram of application request durations. Static-file requests are excluded.

Adds a LatencyHistogram to WorkerMetrics, a MetricValue::Hist variant for IPC transport, histogram exposition in the aggregator, and integration tests covering exposition, counts, static exclusion and the empty case.

Co-authored-by: GitHub Copilot [email protected]

[gi0baro]

A few things:

• the implementation is not measuring the actual e2e latency, as on responses wrapping an async stream the time to send the actual bytes is outside that future block
• not sure why the naming for the is req/request instead of response
• I generally don't allow AI generated code to be merged in Granian, especially if it's added as-is, without the proper refinements (eg: the giant-wall-of-comments is just useless and awful).

[mikmatko]

Thank you for the comments. I've pushed changes.

• the implementation is not measuring the actual e2e latency, as on responses wrapping an async stream the time to send the actual bytes is outside that future block

I believe streaming responses are now correctly tracked.

• not sure why the naming for the is req/request instead of response

OTEL convention, seemingly a "standard" thing for Prometheus metrics.

• I generally don't allow AI generated code to be merged in Granian, especially if it's added as-is, without the proper refinements (eg: the giant-wall-of-comments is just useless and awful).

That's fair :) I've reduced the verbosity of the comments. Personally I think the overly verbose comments are a good thing - it helps a lot when jumping to an unknown codebase. A lot of people read and comprehend natural language faster than code. But yes, it's verbose, maybe even awful, but not useless :)