[Caideyipi]

Summary

• avoid preallocating the complete sliced Load RPC body from the peer-provided original size
• defer rollback cleanup until in-flight Load work finishes and reject new work after rollback is requested
• skip the active-load failure directory when it is nested below a listening directory

Tests

• mvn spotless:apply -pl iotdb-core/datanode
• mvn -pl iotdb-core/datanode "-Dtest=LoadTsFilePieceNodeAssemblerTest,ActiveLoadDirScannerTest,LoadTsFileManagerTest" test (13 tests passed)

[Caideyipi]

Closing in favor of #18936, which preserves all three improvements on dev/1.3. #18698 replaced the global Load manager and EXECUTE/ROLLBACK lifecycle on master with per-DataRegion consensus commands and consensus-aware staged-file retention, so the cleanup/rollback implementation in this PR no longer fits the master Load path. The buffer-allocation and failure-directory scanning improvements remain independently applicable to master, but the legacy cleanup changes should not be restored there just to resolve these conflicts.