* Deployment-level observability to the invoker subsystem, combining
  issues restatedev#4553 and restatedev#4454 into a single change. Lets operators distinguish
  slow user services from internal Restate bottlenecks.

  New metrics:
  - restate.invoker.http_request_duration.seconds (TTFB)
  - restate.invoker.http_total_duration.seconds (total duration)
  - restate.invoker.queue_duration.seconds (internal wait time)
  - restate.invoker.active_invocations (concurrency gauge)
  - restate.invoker.http_status_code.total (non-200 responses)
  - restate.invoker.throttle_balance (token bucket debt)

* Label enrichment on invocation_tasks, task_duration, eager_state_truncated,
  enqueue, and concurrency metrics with partition_id, service_name, and
  deployment_id where applicable.

* ServiceMetrics struct with gauge/counter/histogram helpers.All service
  metric recording goes through the struct methods including TaskMetrics
  struct handles INVOCATION_TASKS lifecycle counters with status/transient
  labels.

* ResponseStream::instrument(metric) builder for HTTP timing, enqueued_at
  timestamp on InvokeCommand for queue duration, and event-driven throttle
  balance recording at slot acquire/release.

* LazyIntern for generic string interpolation including STATUS_CODE_LOOKUP
  for zero-allocation HTTP status code interning.

…v#4559)

Add comprehensive metrics to the networking layer's connection pool via
two scoped structs:

NetworkMetrics (swimlane-scoped, pre-cached per variant):
- pool.connections.active gauge
- connection.pending gauge
- connection.acquisition.duration histogram (cache=hit/miss, result)
- connection.handshake.duration histogram (result=success/error)
- rpc.duration histogram (result=success/error)

ConnectionMetrics (swimlane + peer_name + direction, stored on Connection):
- connection.opened/closed counters (+ legacy connection_created/dropped)
- permit_acquisition.duration histogram
- connection.duration (lifetime) histogram

Key design decisions:
- NetworkMetrics instances pre-built in a static array indexed by Swimlane
- ConnectionMetrics created via NetworkMetrics::connection(peer, direction)
- All labels are &'static str via LazyIntern for peer IDs — zero allocation
- Legacy metrics (connection_created/dropped) preserved with enriched labels
- Pool gauge decremented in deregister() to avoid double-count on drain+drop
- LazyIntern extracted to restate_core::metric_definitions for reuse

Also renames invoker metric constants to follow OTel naming conventions.

… import

Clippy flags new_without_default on LazyIntern<K>::new(). Add the
trivial Default impl. Also remove an unused std::sync::LazyLock import
in the network metric_definitions introduced by the connection pool
instrumentation commit.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Remove redundant DashMap pre-check in LazyIntern::get; entry() handles hits
- Restore original metric name restate.invoker.concurrency_limit to avoid breaking dashboards
- Use "unknown" placeholder for deployment_id label instead of omitting it, keeping Prometheus label sets uniform

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…helper

Remove the #[cfg(test)] EMPTY const from ServiceMetrics and introduce
local test_metrics() helpers that use realistic fake labels ("test",
"test.Service") via ServiceMetrics::new().

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>