metrics/prometheus: rename histogram timings from _ns to _seconds

[squiidz]

When use_histogram_timing is true, timing metrics are recorded in seconds but were still exported under their _ns name. This both misreported the unit and, in a fleet with mixed use_histogram_timing settings, caused the same metric name to be sent as two Prometheus kinds (summary vs histogram). Remote-write targets such as Vector and Mimir reject the whole batch in that case ("multiple metric kinds given for metric name"), dropping data and adding backpressure.

Rewrite a trailing _ns suffix to _seconds in histogram mode so the histogram series (e.g. processor_latency_seconds) is distinct from the summary series (processor_latency_ns) and the unit suffix is correct.

Fixes INC-1095.

[claude]

Commits
LGTM

Review
Small, focused fix that rewrites the _ns suffix to _seconds for timing metrics when use_histogram_timing is enabled, correcting the unit label and keeping histogram series distinct from summary series to avoid Prometheus remote-write "multiple metric kinds" rejections. The rename is correctly scoped to the histogram path only, applied after metric-name validation, and covered by tests for both histogram and summary modes (including inverse NotContains assertions). Docs and changelog are updated consistently.

LGTM

[claude]

---

Commits
LGTM

Review
This PR rewrites a _ns timing-metric suffix to _seconds when use_histogram_timing is enabled, correcting the unit label and keeping the histogram series name distinct from the summary series (avoiding Prometheus remote-write "multiple metric kinds" rejections in mixed fleets).

The change is correct and well-scoped: metric-name validation in NewTimerCtor runs on the original path before the rename, the rename only appends a valid suffix, and the renamed path is used consistently as both the timersHist map key and the histogram Name. Unit tests cover both histogram and summary modes (including negative assertions), and the docs/CHANGELOG are updated.

LGTM

---