feat(cudf): Support Spark string-to-decimal try_cast

[thirtiseven]

Summary

• Add a Spark-compatible GPU implementation of try_cast(VARCHAR AS DECIMAL).
• Preserve HALF_UP rounding, scientific notation, precision and scale checks, whitespace handling, and null-on-invalid semantics for DECIMAL64 and DECIMAL128.
• Route decimal logical types away from the existing string-to-integral path.

Motivation

Velox short decimals use an integral physical type, so the cuDF expression adapter previously classified a string-to-decimal cast as a string-to-integer cast. Long decimals were rejected outright. libcudf's fixed-point parser is not a semantic substitute because it truncates fractional values where Spark rounds HALF_UP.

The conversion kernel is adapted from the Apache-licensed NVIDIA/spark-rapids-jni implementation, with additional validation for whitespace-only strings and incomplete mantissa or exponent forms.

Scope

This PR supports non-ANSI try_cast only. It does not change generic cast behavior, introduce a runtime dependency on RAPIDS JNI, or affect non-decimal cast paths.

Validation

• velox_cudf_spark_string_to_decimal_test: 5/5 passed.
• GPU results are compared directly with the Spark SQL CPU evaluator across short and long decimals, non-zero scales, HALF_UP boundaries, scientific notation, whitespace, invalid input, precision overflow, nulls, and a composed validation predicate.

[github-actions]

Selective Build Plan

Linux release with adapters is running a full build (changes touch velox/experimental/ or velox/external/). See the CI workflows README for what this means.

---

Selective build plan