add basic metrics for NRI framework

[chrishenzie]

Provides necessary hooks for container runtimes to collect basic runtime metrics regarding NRI plugin activity. This improves observability and helps satisfy v1.0 requirements.

Introduces a Metrics interface that allows us to collect plugin counts, invocations, adjustments, and latency information.

To avoid burying state change events under a generic operation name, StateChange invocations append the specific event type using human-readable strings (e.g., "StateChange/CreateContainer").

These hooks are called during plugin invocations and container lifecycle events. Runtimes can provide their own implementation of the Metrics interface (e.g. via a WithMetrics option) to route these data points into their existing telemetry setups.

Ref: #272
Signed-off-by: Chris Henzie chrishenzie@gmail.com
Assisted-by: gemini-cli

[samuelkarp]

Marking as draft per offline conversation with @chrishenzie. Some unstructured thoughts (haven't taken a deep look yet):

• CI is failing
• Bringing the whole Prometheus library stack into the go.mod is a lot
• Clients will have to plug in Prometheus themselves anyway, so a couple alternatives:
  1. Build tags (no_prometheus) for consumers that don't want it, like plugins. But that still ends up in the vendor folder.
  2. A Metrics interface with members like RecordPluginInvocation, RecordPluginAdjustments, etc, and then have runtimes provide an impl tied to their own Prometheus setup as desired (or, could be something really, really simple like a stdout logger)

[samuelkarp]

Looking more, I think a Metrics interface would be helpful. containerd uses https://github.com/docker/go-metrics as a wrapper around the Prometheus libraries anyway, and the interface would allow that to still be used for NRI metrics.

[chrishenzie]

Refactored to introduce a Metrics interface, and provide more fine-grained labels for StateChange invocations.

[samuelkarp]

I expect this will cause some conflicts with #274

[klihub]

@chrishenzie Thank you for picking this up ! With this updated interface-based approach this looks pretty good to me.

[klihub]

@chrishenzie I think the CDO check gets confused by the Assisted-by: being at the very end.

[chrishenzie]

Updated to also collect latency metrics, and refactored test suite to collect metrics via the dummyPlugin.