Introduce CI infrastructure for cuda.parallel

[gevtushenko]

We just introduced cuda.parallel module. This issue is about adding CI infrastructure. Testing is similar to cuda.cooperative module. In terms of dependencies, cuda.parallel module depends on libcu++, CUB, and Thrust, and CCCL/C, as well as gcc 13 compiler.

Tasks

Beta Give feedback

1. Add running cuda.parallel build/test to test_pycuda.sh script/jobs

[gevtushenko]

Ideally, cupy will be the first adopter of the cuda.parallel package. To make cuda.parallel a dependency of cupy, we have to conver similar support matrix as cupy. @leofang can elaborate on what we should test in our CI.

[leofang]

To make cuda.parallel a dependency of cupy, we have to conver similar support matrix as cupy. @leofang can elaborate on what we should test in our CI.

Sorry for delay. CuPy supports the Cartesian product of the following axes (cc: @kmaehashi for vis):

• Python: 3.9-3.12 (the moving window is defined as per NEP-29)
• CUDA: 11.2-11.8, 12.0-12.x (most recent two CUDA major releases that support CUDA minor version compatibility)
• OS: linux-64, linux-aarch64, win-64

For pip installation, pip install cupy-cuda{11,12}x is for CUDA 11.x/12.x. For conda installation, it's conda install -c conda-forge cupy cuda-version={11,12}, see the CuPy Installation Guide.