[SPARK-59949] Record the operator.sdk controller execution histograms with sub-second precision

[sankalpsthakur]

What changes were proposed in this pull request?

This PR records the controller execution histograms of OperatorJosdkMetrics in nanoseconds instead of whole seconds.

• timeControllerExecution measures the duration with Clock.nanoTime() and records it once per execution in both the cluster-wide and the namespaced histogram.
• The histogram names end with nanos, e.g. sparkapplication.sparkappreconciler.reconcile.both.nanos and sparkapplication.sparkappreconciler.reconcile.failure.nanos, so PrometheusPullModelHandler exports them in seconds as operator_sdk_sparkapplication_sparkappreconciler_reconcile_both_seconds etc., like kubernetes_client_http_response_latency_seconds.
• OperatorJosdkMetrics gets a package-private constructor taking a Clock, for tests, following SummingHistogram(Reservoir) of [SPARK-59935] Fix Prometheus _sum and quantiles of histograms and timers #922.
• The migration guide gets an item for the renamed metrics.

Why are the changes needed?

timeControllerExecution updated its histograms with TimeUnit.MILLISECONDS.toSeconds(...), so the elapsed time was truncated to whole seconds and a reconciliation under one second, which most are, recorded 0. The quantiles of these histograms were mostly 0, and their Prometheus _sum (#922) undercounted. This is there since SPARK-48984 (0.1.0).

Does this PR introduce any user-facing change?

Yes. Compared to 1.0.0, the operator.sdk controller execution histograms are renamed with a nanos suffix, e.g. operator_sdk_sparkapplication_sparkappreconciler_reconcile_both_seconds instead of operator_sdk_sparkapplication_sparkappreconciler_reconcile_both in Prometheus, and their values are in seconds with sub-second precision instead of whole seconds. The migration guide is updated.

How was this patch tested?

Updated and new test cases:

• OperatorJosdkMetricsTest.testTimeControllerExecution times the executions with a ManualClock and asserts that a 10 ms success and a 5 ms failure are recorded as 10000000 and 5000000 nanoseconds under the new names, instead of sleeping for a second.
• PrometheusPullModelHandlerTest.testFormatMetricsSnapshotIncludesControllerExecutionHistogramInSeconds asserts that such a histogram is exported as ..._reconcile_both_seconds with 0.01 as its median and _sum.

Nothing was built or run locally; validation is the GitHub Actions run on this PR.

Was this patch authored or co-authored using generative AI tooling?

Generated-by: Claude Code (Claude Opus 5.5)