cuda-bindings: handle CUDA header discovery errors

[yagna-1]

Summary

• Make the cuda-bindings build script report missing or unreadable CUDA headers directly instead of falling through to a bindgen panic.
• Discover CUDA headers in both the standard include/ layout and target-specific redistributable layouts such as targets/x86_64-linux/include/ and targets/sbsa-linux/include/.
• Pass discovered include paths to bindgen and matching target-layout library paths to rustc.

Fixes #87.

Test plan

• cargo fmt --check --all
• cargo test -p cuda-device -p oxide-artifacts -p reserved-oxide-symbols --lib --tests
• cargo test -p cargo-oxide --all-targets on a host without CUDA headers now fails with the direct cuda-bindings: could not find cuda.h... build-script error instead of a bindgen panic.

[fbrv]

I think this is not enough: in sbsa environment this wouldn’t work, even if cuda is installed.
Also it doesn’t take in consideration environments where cuda is not installed in the default path but I think it’s a project decision to support or not default paths (@nihalpasham)

[yagna-1]

Updated to scan CUDA target-layout directories via targets/* instead of hardcoding targets/x86_64-linux, so SBSA layouts such as targets/sbsa-linux are covered as well.

[fbrv]

I think this introduces another problem: if multiple CUDA targets are installed, we may end up mixing dependencies and likely cause build/link errors. We should probably first discover the CUDA host target, and only then look for cuda.h under that target.