feat(metrics): add distribution metric support

[phacops]

Summary

Adds support for distribution metrics to Arroyo's metrics backend interface, alongside the existing increment, gauge, and timing methods.

Distribution metrics let backends record the full distribution of a set of values (e.g. percentiles), which is useful for measuring things like batch sizes or latencies where aggregate statistics matter.

Changes

• arroyo/utils/metrics.py: Add the abstract distribution method to the Metrics protocol and a no-op implementation in DummyMetricsBackend.
• tests/metrics.py: Add a Distribution call type and implement distribution in the testing backend so recorded distribution calls can be asserted in tests.
• tests/utils/test_metrics.py: Add a test covering the testing backend's distribution method.
• docs/source/metrics.rst: Document distribution in the example backend implementation.
• CHANGELOG.md: Add an Unreleased entry.

Notes

This is purely additive to the metrics interface. Backends implementing the Metrics protocol will now need to provide a distribution method.

Testing

• pytest tests/utils/test_metrics.py passes.
• Verified DummyMetricsBackend and the testing backend still satisfy the runtime_checkable Metrics protocol.

🤖 Generated with Claude Code

https://claude.ai/code/session_019ASHzHqVirSvv3vBRYaQPS

---

Generated by Claude Code

[untitaker]

Distribution metrics let backends record the full distribution of a set of values (e.g. percentiles), which is useful for measuring things like batch sizes or latencies where aggregate statistics matter.

okay but we already use timing everywhere, and arroyo isn't prescriptive about how you aggregate those values already. so you could impl timing as a histogram, sampled quantiles, or both, or anything else. is the plan to introduce new metrics that use distribution? right now this dead code.

(don't have access to the claude session so I can't figure out what the bigger picture here is)

[phacops]

With us moving away from veneur and using ddog-statsd, we can now use real distribution metrics instead of the histogram + split into different percentile metrics at veneur.

This PR is adding support for these metrics in Arroyo so we can use them in Snuba consumers https://github.com/getsentry/snuba/blob/10cd401703ce9bd5c838ad38a99a834d4c963957/rust_snuba/src/metrics/statsd.rs#L124-L162