GH-1067: Close cached HDFS FileSystem instances

[xborder]

Summary

Fixes apache#1067.

This change adds explicit HDFS resource cleanup in the same spirit that the dataset JNI layer already does explicit lifecycle cleanup for S3. For S3, Arrow Java registers a shutdown hook that finalizes the native S3 subsystem. For HDFS, the leak happens through cached Hadoop FileSystem instances that keep non-daemon IPC threads alive, so the cleanup needs to happen when FileSystemDatasetFactory is closed.

FileSystemDatasetFactory now retains its input URIs and, on close(), detects HDFS-backed URIs, normalizes and deduplicates the filesystem roots, and best-effort closes the corresponding cached Hadoop FileSystem instances. That releases the lingering IPC threads and allows the JVM to exit normally after HDFS dataset reads.

This PR also adds a regression test using MiniDFSCluster and a forked child JVM to show the behavioral difference:

• without cleanup, the child JVM hangs
• with FileSystemDatasetFactory cleanup, the child JVM exits normally

Testing

mvn -Parrow-jni -pl dataset -Dmaven.gitcommitid.skip=true -Dtest=TestHdfsFileSystemCleanup test

Result:

• Tests run: 2, Failures: 0, Errors: 0, Skipped: 0
• BUILD SUCCESS