[TASK] Post-CUDA 13 driver fixups

[gmarkall]

In getting CI working with the CUDA 13 driver there were a couple of hacks needed, which should be addressed when possible:

• AArch64 / Python 3.11 is tested with CUDA 12.2 instead of CUDA 12.0 as cuda-python 12.9.1 cannot be installed with CUDA 12.0 or 12.1 yet. This change should be reverted when the latest 12.9 bindings can be installed with CUDA 12.0. See https://github.com/NVIDIA/numba-cuda/pull/385/files#diff-cc749e1d52d834b934e6d416112a352bd963b19a7c94fb382be77784044ea0c4R13
• Test binaries are not built when testing the wheel with the ctypes bindings. This is because the build system uses cuda.bindings.nvrtc, which is not installed in that setup. Some way to build the LTOIR objects in this configuration needs to be found. See https://github.com/NVIDIA/numba-cuda/pull/385/files#diff-c6a6ecd17bfcae0e053cf9dce661870baff104ad82042ce6af76ecb15c6eec0f
• There are some nvdisasm issues - nvdisasm from toolkits different to (older than?) the driver version sometimes struggle with the binaries produced by the latest driver / nvJitLink, and I have not been able to identify a pattern. The SASS inspection tests were disabled, but they should be re-enabled and the root cause resolved (if it was not a transient issue with the packaged versions we were getting in CI shortly after the CUDA 13 release). See https://github.com/NVIDIA/numba-cuda/pull/385/files#diff-a9317012eb2763cef1081435b171637317071e1ab252ab82f00bb8026241409b

[kkraus14]

• AArch64 / Python 3.11 is tested with CUDA 12.2 instead of CUDA 12.0 as cuda-python 12.9.1 cannot be installed with CUDA 12.0 or 12.1 yet. This change should be reverted when the latest 12.9 bindings can be installed with CUDA 12.0. See https://github.com/NVIDIA/numba-cuda/pull/385/files#diff-cc749e1d52d834b934e6d416112a352bd963b19a7c94fb382be77784044ea0c4R13

This should be resolved by conda-forge/cuda-python-feedstock#139

[kkraus14]

• Test binaries are not built when testing the wheel with the ctypes bindings. This is because the build system uses cuda.bindings.nvrtc, which is not installed in that setup. Some way to build the LTOIR objects in this configuration needs to be found. See https://github.com/NVIDIA/numba-cuda/pull/385/files#diff-c6a6ecd17bfcae0e053cf9dce661870baff104ad82042ce6af76ecb15c6eec0f

We have decided this is a wontfix because we plan to drop the ctypes bindings in the future.