Fix DML misclassification for statements containing UNION/INTERSECT/EXCEPT

[msrathore-db]

Summary

Fixes #1418.

Statement.execute() was incorrectly returning true (and getUpdateCount() returning -1) for INSERT / UPDATE / DELETE / MERGE statements whose subqueries or CTEs contained UNION, INTERSECT, or EXCEPT. The UNION_PATTERN, INTERSECT_PATTERN, and EXCEPT_PATTERN regexes in DatabricksJdbcConstants are non-anchored (\s+UNION\s+ etc.) and were matched via find() inside shouldReturnResultSet, so the keyword was picked up anywhere in the SQL — even deep inside a subquery of an outer DML, or inside the Databricks column-exclusion form SELECT * EXCEPT (col).

Downstream:

• executeUpdate() threw DatabricksSQLException because shouldReturnResultSet==true triggers the post-execution guard at DatabricksStatement.java:103-108.
• getUpdateCount() returned -1 because executeInternal forces updateCount = -1 when shouldReturnResultSet==true.
• Frameworks that use !execute() as the DML detector (e.g. Slick's sqlu) crashed on perfectly normal DML.
• This was a regression from Simba 2.7.5, which returned false for these inputs.

Fix

Short-circuit shouldReturnResultSet to false when the trimmed query starts with a DML keyword, so the non-anchored set-operator patterns can't fire on subquery content:

if (DML_PREFIX_PATTERN.matcher(trimmedQuery).find()) {
    return false;
}

A separate DML_PREFIX_PATTERN (matching ^(\s*\()*\s*(INSERT|UPDATE|DELETE|MERGE)\s+) is added so the existing INSERT_PATTERN — shared with InsertStatementParser and requiring INSERT INTO — is untouched. This also lets the new guard cover INSERT OVERWRITE ... (called out in the issue).

The existing NonRowcountQueryPrefixes opt-in still wins: it is evaluated before the new short-circuit, so a user who has explicitly set e.g. NonRowcountQueryPrefixes=INSERT still gets ResultSet mode for INSERTs.

Why this is safe

• Every existing top-level UNION / INTERSECT / EXCEPT test case continues to classify as ResultSet via the already-anchored SELECT_PATTERN / WITH_PATTERN / VALUES_PATTERN / FROM_PATTERN. A regression test for (SELECT ...) UNION (SELECT ...) is included.
• Plain DML (INSERT INTO t VALUES (1), UPDATE t SET c=1, DELETE FROM t, MERGE INTO ...) previously returned false because none of the OR-chain patterns matched; it continues to return false now, via the short-circuit.
• INSERT_PATTERN is not modified, so InsertStatementParser / EnableBatchedInserts are unaffected.
• All 3305 tests in jdbc-core pass.

Test plan

• New unit tests in DatabricksStatementTest covering:
  • INSERT ... SELECT ... FROM (... UNION ALL ...)
  • INSERT ... FROM (... INTERSECT ...)
  • INSERT ... FROM (... EXCEPT ...)
  • INSERT ... SELECT * EXCEPT (col) FROM src (Databricks column-exclusion)
  • INSERT OVERWRITE DIRECTORY ... INTERSECT ...
  • UPDATE ... (... UNION ...)
  • DELETE ... (... EXCEPT ...)
  • MERGE INTO ... USING (... UNION ...) ...
  • NonRowcountQueryPrefixes=INSERT opt-in still forces ResultSet mode
  • (SELECT ...) UNION (SELECT ...) still classified as ResultSet (regression guard)
• All 102 existing DatabricksStatementTest cases pass (including top-level UnionQuery / IntersectQuery / ExceptQuery).
• All 52 DatabricksPreparedStatementTest cases pass.
• All 24 InsertStatementParserTest cases pass (batching parser untouched).
• Full mvn test -pl jdbc-core clean: 3305 tests, 0 failures, 0 errors.

NO_CHANGELOG=true