Fluss source: consume Java-client Arrow batches

Goal

Add a StreamFusion Fluss source path that keeps Arrow buffers returned by the Fluss Java client instead of first materializing live-log records as Flink RowData.

This should use Fluss Java APIs for split/offset/checkpoint coordination and direct Arrow polling for the data plane. It should not depend on LogScannerImpl, other @Internal classes, or unpublished locally built Fluss artifacts.

Upstream dependency

Wait until the direct Arrow polling API is public and available in a published Fluss artifact:

• Apache Fluss issue: https://github.com/apache/fluss/issues/4005
• Apache Fluss PR: https://github.com/apache/fluss/pull/4006

The upstream PR exposes LogScanner.pollRecordBatch, preserves the stored PK-table changelog sidecar, and supports non-empty top-level Arrow projections. Fluss 0.9.1 does not contain the batch API; the current 1.0 snapshot exposes only an internal append-only, unprojected path and is not an acceptable production dependency.

Initial admission whitelist

Accelerate only when all applicable contracts have been verified:

• table.log.format = ARROW.
• The live/unbounded log phase can return owned Arrow batches directly.
• Append-only tables, plus primary-key changelogs only when the public API includes aligned per-row change types.
• Full scans or verified non-empty top-level projections.
• Supported Arrow/Flink types and schema evolution.
• Supported client security and configuration options.
• Flink-equivalent split assignment, startup offsets, partition discovery, checkpoint recovery, watermarks, idleness, and metrics.
• Explicit Arrow ownership and StreamFusion off-heap accounting.

Everything outside the whitelist must fall back at planning time to the stock Flink Fluss source. Snapshot/bounded reads may remain on the ordinary Flink row path until Fluss exposes an equivalent verified Arrow snapshot contract.

Acceptance criteria

• Depend on a released Fluss artifact containing the public Arrow polling API.
• Add a planner matcher that implements the whitelist and explains every fallback.
• Preserve append-only and PK changelog semantics, offsets, checkpoint recovery, projection ordering, and schema evolution.
• Prove allocator/version compatibility at the Java-to-native Arrow boundary.
• Add integration and failover tests against the stock Flink Fluss connector as the semantic oracle.
• Benchmark the direct Arrow path against the existing row-materializing path.
• Update docs/connectors/fluss.md with the exact accelerated and fallback circumstances.

Start by checking Apache Fluss issues #4005 and #4006 and confirming that a released artifact provides the public Arrow polling API. Then trace the planner matcher, the stock Flink Fluss connector used as the semantic oracle, and docs/connectors/fluss.md. Done means the whitelist and fallback explanations, integration and failover tests, compatibility checks, benchmark, and documentation updates are complete.