SCHED-316: Add configurable job sources for Slurm exporter

[Uburro]

Problem

The exporter only read jobs from the Slurm controller API. On large clusters where the controller endpoint is overloaded, there was no way to switch the exporter to the accounting API as a fallback.

Solution

Added a JobSource knob with two values: controller (default, behavior unchanged) and accounting (slurmdbd). Only one source is used at a time, no dedupe or fan-out.

When accounting is selected:

• Time window is [now - AccountingJobsLookback, now + 5m]. Lookback defaults to 1h and accepts both Prometheus durations (1d, 1w) and Go durations (1.5h, 2h45m30.5s). The +5m end-time skew tolerates clock drift between slurmrestd, slurmctld, and slurmdbd.
• AccountingJobStates is a list of Slurm job-state strings (PENDING, RUNNING, COMPLETED, ...) forwarded verbatim to the accounting state query parameter. Empty list = no filter. Slurmdbd applies this to the historical states a job held during the window (sacct semantics), not to the current state of returned jobs.
• disable_truncate_usage_time=true, so original start/end timestamps survive the lookback window.
• skip_steps=true, to avoid downloading per-step payloads the exporter never reads.
• Cluster is intentionally left unset; soperator targets a single Slurm cluster per deployment.
• For pending multi-node jobs the accounting API doesn't expose the originally requested node count, so slurm_job_memory_bytes is omitted rather than reported as MemoryPerNode × 1 (which would silently undercount).

CRD fields (jobSource, accountingJobStates, accountingJobsLookback) are all marked EXPERIMENTAL in their godoc / CRD description. accountingJobsLookback has a CEL validation that rejects zero values. Lookback parsing is gated on --job-source=accounting, so a stale env value can't break the default controller path.

Testing

• go build ./..., go vet ./...
• go test ./... (full suite green)
• helm unittest helm/slurm-cluster (89/89)

New unit tests cover source dispatch, accounting state CSV, time-window via testing/synctest, disable_truncate_usage_time and skip_steps params, the dual Prom/Go duration parser, controller-mode tolerance for stale lookback values, and the NodeCount=nil memory-omission behavior.

Release Notes

Feature: Slurm exporter can now collect jobs from the Slurm accounting API (slurmdbd), useful when the controller endpoint is overloaded on large clusters.
Config: New CRD fields jobSource, accountingJobStates, accountingJobsLookback (and matching exporter flags --job-source, --accounting-job-states, --accounting-jobs-lookback). All experimental.

[theyoprst]

I don't get why do we need this. Initially we had a problem that exporter is too slow for big clusters. But this change adds retrieval of jobs from accounting, which will slow it down further.
@Uburro how it will help?

[Uburro]

I don't get why do we need this. Initially we had a problem that exporter is too slow for big clusters. But this change adds retrieval of jobs from accounting, which will slow it down further. @Uburro how it will help?

added the ability to collect metrics from both the controller and accounting, as well as to choose which metrics to collect. For large clusters, you can trade off certain metrics in favor of performance.