fix: [df52] route timestamp timezone mismatches through spark_parquet_convert

[andygrove]

Summary

• Fix incorrect timestamp timezone handling in the schema adapter on the df52 branch
• INT96 Parquet timestamps coerced to Timestamp(us, None) by DataFusion were being routed through Spark's Cast expression when the logical schema expected Timestamp(us, Some("UTC")). Spark's Cast treats None-timezone as TimestampNTZ (local time) and applies a timezone conversion, shifting values by the session timezone offset (e.g., -5h45m for Asia/Kathmandu)
• Route Timestamp -> Timestamp mismatches through CometCastColumnExpr which delegates to spark_parquet_convert, handling this as a metadata-only timezone relabel

Test plan

• Existing test "SortMergeJoin with unsupported key type should fall back to Spark" in CometJoinSuite should now pass
• Existing Rust tests (parquet_roundtrip_int_as_string, parquet_roundtrip_unsigned_int) continue to pass
• CI passes

🤖 Generated with Claude Code

[andygrove]

CI Comparison: PR #3494 vs df52 baseline (PR #3470)

	PR #3470 (df52 baseline)	PR #3494 (with this fix)
Failing CI jobs	34	30
Passing CI jobs	same set + 4 extra failures	same set

Jobs fixed by this PR (fail in #3470, pass in #3494)

• spark-sql-native_datafusion-sql_core-3/spark-3.5.8
• spark-sql-native_datafusion-sql_hive-2/spark-3.5.8
• ubuntu-latest/Spark 3.5, JDK 17, Scala 2.12/native_datafusion [exec]
• ubuntu-latest/Spark 3.5, JDK 17, Scala 2.12/native_datafusion [sql]

Regressions introduced by this PR (pass in #3470, fail in #3494)

None — every failure in #3494 also exists in #3470.

Summary

This PR is a strict improvement over the df52 baseline: 4 fewer failing CI jobs with zero regressions.

---

Note

This comment was generated with the assistance of AI (Claude Code) and should be verified independently.