HBASE-30398 Support configuring the concurrency of off-peak compactions

[Huginn-kio]

HBASE-30398

HBASE-30398 tracks the issue described below.

Summary

HBase currently allows only one off-peak compaction at a time within a RegionServer. This is
enforced by a single AtomicBoolean shared by all stores, so compaction capacity is underutilized
on systems with enough resources.

Root cause

HStore uses one AtomicBoolean for off-peak compaction admission. A store can enter the off-peak
path only when it changes the flag from false to true; the flag is cleared after selection
fails or the compaction finishes. As a result, all stores contend for one permit.

Fix

• Replace the single boolean gate with a shared semaphore-backed tracker.
• Add hbase.hstore.offpeak.compaction.concurrency, defaulting to 1, to control the number of
  concurrent off-peak compactions.
• Register the tracker with configuration observers so the permit count follows runtime
  configuration changes.
• Add unit coverage for the configured concurrency and updates to the setting.

Testing

• mvn -pl hbase-server -am -Dtest=TestOffPeakCompactionTracker -Dsurefire.failIfNoSpecifiedTests=false -DskipITs test
• mvn -pl hbase-server -am -Dtest=TestOffPeakCompactionTracker,TestOffPeakHours,TestCurrentHourProvider,TestCloseChecker -Dsurefire.failIfNoSpecifiedTests=false -DskipITs test
• mvn -pl hbase-server -am -DskipTests package
• mvn -pl hbase-server -DskipTests checkstyle:check -Dcheckstyle.includes='**/OffPeakCompactionTracker.java,**/TestOffPeakCompactionTracker.java'
• mvn -pl hbase-server -DskipTests apache-rat:check spotbugs:check

The focused tests passed with 2 tests and the adjacent regression slice passed with 7 tests.

Compatibility

The setting is optional and defaults to 1, preserving the current behavior. This change does not
alter public APIs, wire formats, or persisted state.