Prevent MERGE replay after ambiguous commit

[KirillKurdyukov]

Pull request type

• Bug fix

What is the current behavior?

If Connection.commit() throws a normally retryable YDB error, YdbMergeSink treats the transaction as uncommitted and can replay the buffered MERGE on a new connection even though the commit outcome is ambiguous.

What is the new behavior?

• Mark the commit boundary immediately before invoking Connection.commit().
• Never replay a MERGE after commit has been invoked, regardless of retry classification.
• Preserve the original SQL exception as the cause of the unknown-outcome error.
• Add one focused fault-injection unit test that executes a batch, throws retryable ABORTED from commit, and verifies one connection, batch execution, and commit invocation.

Validation

• JDK 25 compile passed: mvn -f ydb-trino-adapter/pom.xml -DskipTests compile.
• Focused unit passed: mvn -f ydb-trino-adapter/pom.xml -Dtest=TestYdbMergeSink test — 1 run, 0 failures, 0 errors, 0 skipped.
• Independent review found no findings on head 599332520f9ec08edf57634fa18266a1c3fda83a.
• Local Docker-backed inherited testMerge*, smoke testMerge, and full suite were not run because the existing Colima runtime is broken (colima status: error retrieving current runtime: empty value; Docker socket unavailable). The full GitHub Actions suite is required for this PR.