CI: Prevent/kill Jenkins k8s agent orphans and unschedulable agent pods

[michaelsembwever]

node pools can get stuck at max desired size, wasting a lot of resources and $$$
Either the instanceCap goes beyond the node group sizes, and it thrashes, or (for example) the controller fails its liveness probe, gets SIGKILLed, then orphans agent pods forever. Manually trying to scale down the node groups won't help the as the jenkins keeps trying to scale up again…

jenkins-deployment.yaml:

• enable agent.garbageCollection, so a controller restart no longer leaks agent pods; activeDeadlineSeconds cannot tell an orphan from a legitimate 6h microbench build
• cap agent-dind-large at the 160 nodes its pool holds. At 200 the surplus pods could never be scheduled, so they churned and drove both large node groups to maxSize
• loosen the controller probes, so a stall no longer takes every running build with it
• log gc and safepoints to the PVC, to make the next stall attributable

run-ci:

• refuse to deploy an instanceCap above what the node pools can schedule, read from the cluster-autoscaler status configmap
• fail rather than fall back to origin when a branch tracks no remote, since origin carries no dev branches; suggest push.autoSetupRemote
• helm --timeout 10m, as --wait's 5m default is now shorter than the startupProbe's boot budget

jenkins-check.yaml

• fix helm installation

patch by Mick Semb Wever; reviewed by TODO for CASSANDRA-21576

Assisted-by: Claude Code:claude-opus-5