fix: Respect native Orin CUDA architecture and CuTe link requirements

[suharvest]

What does this PR do?

Fixes the two CMake propagation issues reported in #117 that block a native
Jetson Orin build with JetPack 6 / CUDA 12.6:

1. Capture an explicitly supplied CMAKE_CUDA_ARCHITECTURES before
   project() enables CUDA, so the project default no longer replaces
   -DCMAKE_CUDA_ARCHITECTURES=87.
2. For CuTe DSL link targets that are static libraries, propagate the cudart
   shim with PUBLIC scope and the CUDA 12.0–12.6
   -Wl,--wrap=_cudaLaunchKernelEx option with INTERFACE scope. The existing
   behavior for non-static targets is unchanged.

This branch has been rebuilt on the current v0.9.1 main
(7f061f21f0a581ba234a1e233c9315b89d8e47d6). The v0.8-era patch was not
blindly rebased because cmake/CuteDsl.cmake changed substantially in v0.9.1.

Scope

• Two CMake files only.
• No runtime, model, kernel, or API changes.
• CUDA_DRIVER_LIB remains PRIVATE; this PR does not expand its transitive
  link scope.
• Default x86 architecture selection and the AArch64 toolchain path retain
  their existing behavior.

Validation

• Focused CMake checks cover explicit SM87, default native, and AArch64
  toolchain architecture selection.
• A focused harness including the updated cmake/CuteDsl.cmake verifies that a
  static consumer exposes the CuTe archive and shim through
  INTERFACE_LINK_LIBRARIES, and exposes --wrap=_cudaLaunchKernelEx through
  INTERFACE_LINK_OPTIONS.
• The same harness verifies that CUDA_DRIVER_LIB remains private.
• git diff --check passes.
• A merge-tree check against current main is clean.

A native Orin build and the repository-wide pre-commit suite were not rerun on
the machine used for this v0.9.1 rebase because it has no CUDA/Jetson toolchain
or repository pre-commit environment. CI or Jetson validation is therefore
still requested.

Closes #117.

[rickwalder]

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