PHOENIX-7793 Addendum: Replication Log writer improvements

[tkhurana]

Summary

• On-demand rotation uses a CountDownLatch instead of blind Thread.sleep, waking the retry immediately when a fresh writer is staged
• Every retry gets a fresh writer (rotation requested on first failure, not just the 2nd)
• Rotation size clamped to HDFS block size to prevent single-file multi-block writes
• LogFileFormatWriter syncs header on construction, forcing HDFS block allocation on the rotation thread
• LogRotationTask catches Throwable (not just IOException) to prevent RuntimeException from silently killing the ScheduledExecutorService
• Default sync retries reduced from 4 to 1 (2 total attempts), matching the "fast fail to SAF" design
• calculateSyncTimeout() simplified to derive from hbase.regionserver.wal.sync.timeout + ZK session
• Abort on double-failure: if both SYNC and SAF writes fail, the RS aborts so preWALRestore can re-ship orphaned edits
• PhoenixWALSyncTimeoutException wraps timeout errors for clearer diagnostics
• New syncToSafTransitions metric counter
• Unified ReplicationLog.close(boolean graceful) replaces separate close()/closeOnError() — writer closes submitted async with bounded 10s await
• LogFileWriter.close() uses AtomicBoolean CAS to prevent concurrent double-close
• onExit(graceful) in mode implementations simplified to pass the flag through directly

Test plan

• ReplicationLogGroupTest (39 tests) — validates retry semantics, rotation, abort, timeout, mode transitions
• LogFileWriterSyncTest — validates sync call ordering after header-sync-on-init change
• Run tests in loop to confirm no flaky timing issues
• Integration tests (ReplicationLogGroupIT) for end-to-end with mini-cluster