feat(service): Add span instrumentation to backends

[jan-auer]

Backend operations now appear as real child spans in Sentry, turning the existing per-request and per-task transactions into full trace waterfalls. They were empty before because the sentry-tracing layer only converts INFO+ spans by default; the logging layer now forwards DEBUG+ spans (TRACE still ignored) and the backend spans moved to debug.

GCS/S3 HTTP calls and BigTable RPCs each get a per-attempt span covering the real request duration (method/URL/status, or action/gRPC status), and the tiered, changelog, and per-backend operations are annotated so the waterfall reads top-down. Spans carry small primitive fields (object id, range, part numbers, URLs); payloads, streams, and metadata are excluded.

Observability-only — no behavior changes and console logging is untouched.

Closes FS-412

[linear-code]

FS-412

[lcian]

Looks really good!

The span op/name in Sentry is a bit verbose (as it always was).
This can be sometimes useful as it gives you the module where the call came from (useful to tell e.g. which backend), so I might just have to get used to it.

[jan-auer]

For a moment I was considering to strip the crate names and just leave the last module, but then I came back to the full module. Maybe this is something the UI could collapse in the future.

We can for sure improve the span name, however. It currently just duplicates the function name, but there's more information we could put into the span like instrumentation in other environments does.