HBASE-30411 Support regex based RS membership assignment to RS group - #8733

No functional change; reflows comment line breaks to match the
project's formatter output.

…ng the way

RSGroupInfoManagerImpl: add handleServerEvent(), called synchronously from
serverAdded/serverRemoved. A recompute that only affects 'default'
membership now applies immediately in-memory instead of waking the
background ServerEventsListenerThread; only a regex-governed group's
membership change (which must be persisted to hbase:rsgroup/ZK) hands
off to that thread. Also fix the thread's run loop to wait for a
serverChanged() signal before recomputing, instead of recomputing once
unconditionally on start and only waiting after.

TestRSGroupsBase: strengthen addGroup/removeGroup assertions (ported
from branch-2) to verify servers actually land where expected, and
switch the table-membership check to ADMIN.getRSGroup(TableName)
instead of defaultInfo.getTables() -- MasterRpcServices.getRSGroupInfo
(the live RPC backing ADMIN.getRSGroup(String)) never resolves table
membership from TableDescriptor attributes, unlike listRSGroups() or
getRSGroup(TableName).

TestRegexBasedRSGroupMembership: drop two triggerListenerCycle() calls
(in afterMethod()'s dead hadLeftoverRegex guard, and in clearRegex())
that forced a synthetic server event to mask a regex-governed group's
membership from going stale after a config change. That staleness is
expected -- recompute is event-driven by design, matching branch-2's
semantics -- and removeGroup()/VerifyingRSGroupAdmin#verify() already
tolerate it correctly without help. Also fix several comments left
over from the branch-2 port referencing RSGroupAdminServer and
moveTableRegionsToGroup, neither of which exist on master; point them
at the real master-side methods (RSGroupInfoManagerImpl#setRSGroup /
moveTablesAndWait / moveServers, MasterRpcServices).

Ported INFO-level log lines from branch-2's original regex-based
RSGroup membership feature commits that were missing from master's
adapted implementation: config resolution, hostname matching,
server-to-group resolution, refresh entry/exit, and
validation entry/exit. Also added a server count to the existing
"Updated auto-managed RSGroup servers" log, matching branch-2's
equivalent message.

RSGroupInfoManagerImpl#updateAutoManagedRSGroupServers() used to take a
precomputed assignments map as a parameter. ServerEventsListenerThread#run()
computed that map -- and the comparison against live group state used to
decide whether an update was even needed -- before acquiring the enclosing
instance's lock, the same lock the method itself (and every other mutator)
relies on to serialize access to holder.groupName2Group. A concurrent
admin-driven change (moveServers, flushConfig, etc.) landing in that window
would be silently clobbered by the stale assignments the background thread
had already computed. Make the method self-contained and parameterless: it
now reads holder.groupName2Group, recomputes the auto-managed assignment,
compares against current state, and applies the update, all under its own
synchronized block, so the whole decide-and-apply sequence is atomic.

Separately, ServerEventsListenerThread's own wait/notify signal used a
boolean changed flag that was only cleared after calling
updateAutoManagedRSGroupServers() -- which can block on I/O (hbase:rsgroup
table and ZK writes via flushConfig). A new server event arriving while that
call was in flight would set changed = true, but the subsequent
unconditional changed = false once the call returned would wipe that signal
with nothing left to wake the thread for it. Replace the boolean with an int
eventCount: serverChanged() increments it, and the run loop decrements it by
exactly one per completed pass, so a surplus signal forces an extra
(idempotent) pass instead of being lost.

Also add INFO-level logging for serverAdded/serverRemoved callbacks,
matching the logging already present on the other paths through this class.

TestRegexBasedRSGroupMembership: fix a ~3%-flaky crash-target assumption in
testUngracefulCrashOfRegexMemberReassignsRegionsOnlyToOtherGroupMembers --
retainAssignment's random fallback (BaseLoadBalancer#randomAssignment)
places each region independently, so it's possible for all 6 of this test's
regions to land on sn2 alone instead of being split with sn1. Crash whichever
server actually ended up holding regions, instead of hardcoding sn1, so
ServerCrashProcedure always has something to reassign and
roundRobinAssignment is guaranteed to fire.

…ckground thread

Defensive guard around the eventCount-- in ServerEventsListenerThread.run();
the decrement already only runs once the while-loop confirms eventCount > 0,
so this is a no-op in practice but makes that invariant explicit.

…ity.mdx

The refguide build runs prettier --check on the docs and failed on the two
callouts added for regex-based RSGroup membership.

- Say that hbase.rsgroup.regex.* changes are applied by restarting all
  HMasters (recommended) or, after a reload, only at the next server event
- Update the regex group removal and RegionServer Grouping removal steps to
  restart the HMasters instead of reloading
- Document that adding or changing a regex drops unmatched servers later
- Note that add_rsgroup and rename do not apply a regex