Make CUDA optional at build time to enable Bioconductor release and CPU-only installation

[suzannejin]

Problem

Currently, the package requires CUDA to be present at build time — CMakeLists.txt issues a FATAL_ERROR if CUDA is not found (line 58). This means users without a GPU or CUDA toolkit cannot install the package at all.

Current coupling points

The CPU and GPU code are entangled at four levels, even though the implementations themselves (dispatch/cpu/ and dispatch/cuda/) are already cleanly separated:

Layer	File(s)	Coupling
CMake	CMakeLists.txt:58	FATAL_ERROR if CUDA not found
Rcpp wrappers	src/backend.cpp, src/lrv.cpp, etc. (8 files)	Each #includes the .cuh CUDA dispatch header
Shared header	inst/include/propr/context.h	#include <cuda_runtime.h>, defines propr_context with cudaStream_t
Link step	src/CMakeLists.txt:17	Hard links to CUDA::cublas CUDA::cudart

Deployment targets to support

Allow two install modes: 1) standard mode without cuda, and 2) gpu support with cuda implementations.

Note that it should be Bioconductor compliant (we aim to provide this R package through Bioconductor #81),
and also Bioconda/Biocontainers/conda-forge compliant, so that we can later create official container to use this package on nf-core (check links 1).