#28296 Use CUDART_VERSION reduction compatibility in GQA attention

Description

Update /home/runner/work/onnxruntime/onnxruntime/onnxruntime/contrib_ops/cuda/bert/gqa_unfused_attention.cu to match the existing CUDA attention compatibility pattern used elsewhere in the repo.

• Replace the local reduction functors with the established CUDART_VERSION >= 12090 guards.
• Use ::cuda::maximum() and ::cuda::std::plus() for CUDA 12.9+.
• Keep cub::Max() and cub::Sum() as the fallback for older toolkits.

Motivation and Context

This keeps the GQA unfused attention kernel consistent with nearby CUDA attention code and avoids the CUDA 12.9+ deprecation issue around the old CUB reduction functors while preserving compatibility with older CUDA toolkits.

Validation:

• git diff --check
• Code review validation: no comments
• CodeQL validation: no analyzable language changes detected