[java] Support Arrow batch writes across allocator roots

[jianguotian]

Purpose

MosaicWriter.write(VectorSchemaRoot) exports incoming Arrow batches through the Arrow C Data interface. Arrow 15's Data.exportVectorSchemaRoot reloads the fields into a temporary root struct, which cannot reload buffers owned by an independent allocator root into the writer allocator.

Support those batches without moving or reassociating their buffers: keep the existing same-root path, export cross-root fields directly, and retain temporary Arrow C Data metadata under the writer allocator. Make failed cross-root exports rollback-safe so an allocator OOM or partial child export does not retain input buffers.

Changes

• Recursively validate that all top-level and nested input vectors share one input allocator root.
• Keep same-root batches on the existing Data.exportVectorSchemaRoot path.
• Export cross-root schemas and field arrays directly while charging temporary C Data metadata to the writer allocator.
• Release all successfully exported children and temporary metadata when a later cross-root export step fails.
• Document the synchronous batch lifetime, allocator-root requirement, and writer-owned metadata accounting.
• Keep the native writer API, Rust codec, and on-disk format unchanged.

Tests

• JDK 8: 55 tests passed, 0 failed, 0 errors, 0 skipped.
• JDK 11: 55 tests passed, 0 failed, 0 errors, 0 skipped.
• Covered independent roots, same-root accounting, sequential roots, nested mixed-root validation, writer allocator OOM retry, and partial child-export rollback.
• cargo fmt --all -- --check passed.
• git diff --check passed.