fix: surface SparkArithmeticException(DIVIDE_BY_ZERO) for divide-by-zero in dispatched ScalaUDF path

[0lai0]

Which issue does this PR close?

Closes #4517 .

Rationale for this change

DataFusion 53+ wraps errors with .context(...), producing DataFusionError::Context(_, Box<inner>).

Previous handling mostly matched top-level DataFusionError::External(...), so Context-wrapped Spark errors could bypass typed conversion and leak as CometNativeException instead of Spark’s SparkArithmeticException(DIVIDE_BY_ZERO).

What changes are included in this PR?

• native/core/src/execution/expressions/arithmetic.rs

  • CheckedBinaryExpr::evaluate now matches generic Err(e) (when query context exists), then recursively extracts SparkError via extract_spark_error.
  • Added recursive unwrapping for DataFusionError::Context and nested External.
  • If no Spark-typed error is found, returns Err(e) as-is (instead of re-wrapping as External).

• native/jni-bridge/src/errors.rs

  • Added SparkPayload and extract_spark_payload to recursively unwrap Context / nested External.
  • throw_exception now handles all CometError::DataFusion { source } through extract_spark_payload, covering:
    • JavaException passthrough
    • SparkErrorWithContext JSON throw path
    • bare SparkError JSON throw path
    • fallback generic exception path when no Spark payload is found

• spark/src/test/scala/org/apache/comet/CometCodegenSuite.scala

  • Added regression test: divide-by-zero through dispatched ScalaUDF surfaces SparkArithmeticException (#4517).
  • Enables ANSI mode, runs SELECT 1 / identity_int(a) over data containing zero, and validates exception propagation semantics.

How are these changes tested?

./mvnw test -Dtest=none -Dsuites=org.apache.comet.CometCodegenSuite

[andygrove]

Thanks @0lai0. Can you revert the Spark diff changes made in #4514 as part of this PR?

[0lai0]

Thanks @andygrove. Just to confirm, should I revert both the #4517 workaround (SQLQueryTestSuite DIVIDE_BY_ZERO) and the #4516 workaround (SQLQuerySuite UDF call-count) from dev/diffs?

The #4517 one is clearly no longer needed with this fix, but wanted to check on #4516 since that issue is still open.

[0lai0]

I've reverted the #4514 diff changes. After the revert, udf/postgreSQL/udf-select_having.sql is now failing:
SELECT 1 AS one FROM test_having WHERE 1/udf(a) = 1 HAVING 1 < 2

External: Compute error: Error evaluating filter predicate: External(SparkErrorWithContext { error: DivideByZero, ... })

This PR fixes the error type for the project/aggregate path (e.g. SELECT 1/udf(a) FROM t), but this query goes through filter predicate evaluation instead. The typed SparkErrorWithContext appears to get stringified before it can be unwrapped by the JNI bridge, so it still surfaces as CometNativeException rather than SparkArithmeticException.

Should I extend this PR to cover the filter path too, or file a separate issue and keep this test excluded for now?