[VL] Support type widening in Parquet reader (SPARK-40876)

[baibaichen]

Labels: enhancement, VELOX

---

Description

Enable the GlutenParquetTypeWideningSuite test suite for Spark 4.0 and 4.1, which validates Parquet type widening support (SPARK-40876).

Background

GlutenParquetTypeWideningSuite has 84 tests covering two types of Parquet type conversions:

1. Physical→Logical type restoration: Reading int32 + INT(8) as TINYINT (safe, writer guarantees value range)
2. Schema evolution widening: Reading old IntegerType data as LongType, DoubleType, or DecimalType (Spark 4.0 feature)

Currently the suite is disabled with 74 out of 84 tests failing. The failures fall into four categories:

Category	Count	Issue	Fix
A	13	Velox doesn't support INT→DOUBLE/REAL/DECIMAL widening	Velox C++ convertType() extension
B	29	Exception type mismatch + no Decimal precision check	Exception translation + C++ precision check
C	31	Parquet V2 encoding assertions + Decimal conversion limits	Disable native writer + test overrides + Velox C++
D	1	parquet-mr only decimal narrowing overflow→null	Exclude (cannot reproduce with native reader)

Plan

This will be addressed in 3 PRs:

1. PR 1 — Exception translation: Add translateException() to convert Velox type errors to SchemaColumnConvertNotSupportedException. Enable the suite with appropriate excludes/overrides for tests that pass without C++ changes.

2. PR 2 — SPARK-18108 + Revert OAP: Fix partition column type conflicts. Import upstream Velox PR #15173.

3. PR 3 — Type widening implementation: Velox C++ changes for INT→DOUBLE/REAL/DECIMAL and Decimal→Decimal widening. Requires upstream Velox PR first, then enable remaining tests.

Test Results (Target)

	Spark 4.0	Spark 4.1
✅ Passed	46	46
🟢 Override (passed)	35	35
❌ Excluded	3	3
Total	84	84

Sub-issue of #11550.

This issue was written with the assistance of AI.