Add cross-build support to cuda-toolkit 12.9

[danimtb]

Summary

• Add Linux cross-building support to cuda-toolkit 12.9, matching the 13.x model: patch nvcc.profile to use CONAN_CUDA_TARGET_DIR, package per-arch targets/, and map armv8 to linux-sbsa.
• Reject unsupported arches early (e.g. Windows armv8) instead of failing with a redistrib KeyError.
• Extend CI to create the 12.9 armv8 package on ubuntu-latest; skip test_package on cross creates unless a host compiler can target that arch (cross-test-ref).

arm_target (12.9 only)

CUDA 12.9 still ships separate Arm redistributables (linux-sbsa vs linux-aarch64 / Jetson). This PR adds an arm_target option (sbsa | jetson, default sbsa) so consumers can pick which one to package.

CUDA 13.x redistributables no longer provide linux-aarch64 — only linux-sbsa (Jetson Orin pieces such as libcudla / cuda_compat live under that platform where available). There is nothing to select on 13.x, so armv8 always maps to SBSA and this option is intentionally not added to those recipes.

Test plan

• Native conan create for 12.9 on Linux and Windows
• Cross conan create … -s arch=armv8 on Linux downloads linux-sbsa and lays out targets/sbsa-linux
• Confirm CI native jobs still run test_package; cross create uses --test-folder=""