CI: Take the report merge off the test fan-out, stop the agent pod churn, report split balance

[michaelsembwever]

Three faults measured on a Kubernetes clone of this pipeline.

Jenkinsfile:

• generateTestReports moves from cassandra-medium, the label of the ~460 test splits it queued behind, to cassandra-report: 109 minutes of a 163 minute build, for ~3 minutes of work
• one ant junitreport per target, three at a time under an explicit -Xmx
• size every xargs -P and xz -T from the container's cgroup quota. nproc answers for the node: 8 against a 2 cpu limit
• record each cell's duration and longest suites, and archive them
• test-burn takes 5 splits, which separates its three long classes
• remove testDataPublishers as they were oom-ing ci-cassandra.a.o's controller https://the-asf.slack.com/archives/CK23JSY2K/p1787841152008959?thread_ts=1787836602.176469&cid=CK23JSY2K , and it was noise because of Every test is annotated once per junit() call when a build publishes results more than once jenkinsci/test-stability-plugin#11 (we use butler instead)

jenkins-deployment.yaml:

• slaveConnectTimeout 600 and waitForPodSec 900, the two deadlines on one provisioning attempt. 30 seconds met only a warm node, and churned 1,110 pods in one 24 minute window with 0 agents connected and 483 KubernetesProvisioningLimits ... went below zero warnings
• idleMinutes 5, a pod holding a whole node
• the four instanceCaps to the demand a pre-commit run measured; containerCap to their sum, so no pool's share is decided by which holds its agents longest, and maxRequestsPerHostStr with it
• add the agent-dind-report template, on a node pool of its own

cell_balance.py:

• new. Per target, how evenly its splits divided and what the worst cell ran, to the console and ci_summary.html. Splits stay round-robin, so the lever is the split count
• guarded in generate-ci-summary.sh, never failing a build

run-ci-test.py:

• the fixture's node group maximums follow the raised caps

[michaelsembwever]

squashed and message written.