Fix native Jetson Orin (sm_87) build: respect explicit CUDA arch + propagate CuTe DSL --wrap/shim to static-lib consumers

[suharvest]

What does this PR do?

Type of change: Bug fix

Overview: Fixes two build-system issues that block a native Jetson Orin (sm_87, JetPack 6 / CUDA 12.6) build (closes #117):

1. CMakeLists.txt — respect an explicit -DCMAKE_CUDA_ARCHITECTURES.
   The project default (80;86;89[;100a;120]) was applied whenever AARCH64_BUILD was undefined, overriding an explicit value on a native Orin build. Because enable_language(CUDA) (CMP0104 NEW, the default at our cmake_minimum 3.20) already initializes CMAKE_CUDA_ARCHITECTURES to a compiler default, the user's intent is now captured before project() runs, and the project default is applied only when the user did not pass one. Regular x86 builds are unchanged; explicit selections are preserved.

2. cmake/CuteDsl.cmake — propagate the cudart shim and --wrap to static-lib consumers.
   cute_dsl_setup() attached the cutedsl cudart shim and the CUDA<12.8 -Wl,--wrap=_cudaLaunchKernelEx option with PRIVATE. A static archive performs no link step, so for STATIC_LIBRARY link targets neither reached the consuming executable, breaking the CuTe DSL kernels on JetPack 6 / CUDA 12.6. The shim is now propagated via PUBLIC and the --wrap option via INTERFACE for static-library targets; PRIVATE is retained for shared/other targets.

Usage

Build-system fix; no API change. Native Jetson Orin build:

cmake -B build -DCMAKE_BUILD_TYPE=Release \
  -DEMBEDDED_TARGET=jetson-orin \
  -DCMAKE_CUDA_ARCHITECTURES=87 \
  -DTRT_PACKAGE_DIR=/path/to/TensorRT
cmake --build build -j

🚀 Pull Request Checklist

✅ Pre-commit Checks

• cmake-format (v0.6.13) and codespell run clean on the changed files. (Full pre-commit suite not run locally; the change is limited to two CMake files.)

🧪 Tests

• No automated test covers the cross/Jetson build configuration. Verified manually: a native Orin (sm_87) configure now honors -DCMAKE_CUDA_ARCHITECTURES=87, and the CuTe DSL static-lib consumer links under CUDA 12.6.

📄 Documentation

• No documentation change required (jetson-orin is already a documented EMBEDDED_TARGET).

⚙️ Compatibility

• Backward compatible — x86/datacenter default builds and the aarch64 cross toolchain are unaffected.

Additional Information

Closes #117.

[rickwalder]

+1, this fixes jetson AGX orin build for me (jp 6.2.1)