[platinumhamburg]

Purpose

Part of #4222. Implements the server-side transaction protocol tracked by #4224, following #4242.

Bulk loading an empty primary-key table or physical partition must keep the target inaccessible until every imported bucket is installed and its assigned replicas are ready. The Coordinator now owns a recoverable Begin / GetInProgress / Commit / Abort lifecycle through internal RPCs.

Brief change log

• Persist target ownership and the commit decision; fence ordinary access and stale snapshot or remote-log commits while loading.
• Validate complete bucket coverage and adopt ordinary KV snapshots before atomically activating the target.
• Reconcile Coordinator failover, TabletServer restart, leadership and assignment changes; retain terminal results and expire transaction metadata.
• Track leader activation by notification request and advertise server protocol capabilities.
• Expose active transaction references for the subsequent orphan-cleanup integration. The public SDK, scoped build authorization and Flink integration follow separately.

Tests

• Java 11: 82 focused tests passed across RPC API/error compatibility, ZooKeeper metadata, Coordinator event processing and BulkLoad transaction recovery.
• BulkLoadServerITCase: all 12 E2E cases passed, covering import and continued online writes, abort, target fencing, Coordinator/leader loss, TabletServer restart, partition bucket layouts and terminal-result retries.
• Maven reactor test runs passed Spotless and Checkstyle checks.

API and Format

Adds internal Coordinator RPCs and version 1 of NotifyLeaderAndIsr. Adds optional bulk-load ownership/state and protocol-capability fields to registration metadata, plus persistent bulk-load transaction records. Begin rejects servers that do not advertise the required capability.

Documentation

The public SDK and engine-facing usage documentation are tracked by the remaining tasks under #4222.