separate cpu/cuda modules

[casperdcl]

current

• cuvec.cuvec: at compile time, use cudaMallocManaged/cudaFree if available, otherwise use malloc/free
  • at runtime, people with NVIDIA GPUs are forced to allocate memory on their device

proposed

• cuvec.cuvec_cpu: uses malloc/free
• cuvec.cuvec_cuda: if available, use cudaMallocManaged/cudaFree, otherwise ImportError
• cuvec.cuvec: point to one of the above