cuda: use target-specific paths under CUDA Toolkit on Linux

[MaxandreOgeret]

fix #14595

This PR updates the cuda dependency to use target-specific include and library paths, such as /usr/local/cuda/targets/aarch64-linux/lib and /usr/local/cuda/targets/aarch64-linux/include, instead of defaulting to the host's /usr/local/cuda/lib64.

When cross-compiling, Meson currently:

• Detects the CUDA library directory using the build machine's architecture, not the target's.
• Hardcodes lib64 for aarch64, even though NVIDIA's toolchain uses lib under the target-specific path.
• Ignores the targets/<arch>-<os> layout entirely, making it impossible to use
  CUDA_PATH=/usr/local/cuda/targets/aarch64-linux as a workaround.

This causes incorrect link paths and build failures when targeting non-x86_64 platforms.

Fix:

• Use for_machine to determine the target architecture, not the host.
• On Linux, always use lib under the target path, matching NVIDIA's layout.
• Introduce self.target_path to resolve targets/- subdirectories for both include and lib paths.
• Preserve existing behavior on Windows and macOS.

This fix only applies to Linux.

Tested with jetpack cuda 12.6 & jetpack cuda 10.2
I cannot test it on Windows or other.

Todo:

• Add test for cuda cross