Add conditional CUDA architecture 120 support based on CUDA version

Fix compilation errors on CUDA 12.6 and earlier versions by making
Blackwell architecture (sm_120) conditional. The architecture is now
only included when CUDA Toolkit version is 12.8 or later.

Changes:
- Move find_package(CUDAToolkit) before CMAKE_CUDA_ARCHITECTURES setup
  to access version variables
- Add conditional logic to append architecture 120 only for CUDA 12.8+
- Add informative status messages showing whether Blackwell support
  is enabled
- Maintain backward compatibility with CUDA 11.2 through 12.7
- Remove duplicate find_package(CUDAToolkit) call

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>