Add prometheus-adapter rule generation for nrtsearch GPU autoscaling

[vim345]

Summary

Teaches setup_prometheus_adapter_config.py to read nrtsearch CRD manifests and generate prometheus-adapter rules for GPU-based autoscaling.

Context

The nrtsearch-operator manages its own HPA (not through PaaSTA's autoscaling framework), so the existing get_rules_for_service_instance path doesn't apply. However, the operator's HPA still needs the prometheus-adapter to serve its custom metric.

This adds a new function get_rules_for_nrtsearch_gpu() that:

1. Reads {soa_dir}/nrtsearch/nrtsearchserviceeks-{cluster}.yaml
2. Finds any serverSet with targetGpuUtilization > 0
3. Generates an adapter rule that joins DCGM_FI_DEV_GPU_UTIL with kube_pod_labels to associate GPU utilization with nrtsearch pods
4. Divides by the target so the HPA sees 1.0 when at target utilization

The generated metric name follows the existing convention: {deployment-name}-arbitrary_promql-prom.

Test plan

• New unit tests: rule generation, missing file handling, no-gpu-config skip

[nemacysts]

i think we can probably do this without having to add a special case function for nrtsearch

we'd want to update K8S_INSTANCE_TYPE_CLASSES so that

paasta/paasta_tools/setup_prometheus_adapter_config.py

Line 971 in 85e76d7

for instance_type_class in K8S_INSTANCE_TYPE_CLASSES:

covers nrtsearch configs

then, we probably need to figure out some typing shenanigans for instance_config in get_rules_for_service_instance(), but then as long as we implement/have implemented all the instance_config.FOO() methods we're calling here, we might not need to do much else re: making this work for nrtsearch?

[nemacysts]

especially since serverSets should only have 2 items and only the non-primary is autoscalable, right? ('cause then we can essentially have any of the methods that look up autoscaling data just do that for the non-primary :D)

[vim345]

When series_query is provided in the MetricsProviderDict, the create_instance_arbitrary_promql_scaling_rule path skips the label_replace() wrapping and uses the raw metrics_query/series_query directly. Is that sufficient for the prometheus-adapter to correctly associate the metric with the target deployment, or do we still need the label_replace for deployment/namespace labels in the metrics_query?

[nemacysts]

aren't these going to be in paasta-nrtsearchservices?