Make CuPy an optional dependency

[leofang]

We are building fundamental libraries for all Python GPU libraries and frameworks such as CuPy, PyTorch, etc to depend on. So it is important that we do not make any of them a required dependency to avoid circular import and difficulty in updating CUDA support. We can, however, use them in code samples and tests as usual, as long as proper import guards are added. All CUDA driver/runtime/compilation capabilities should switch to use cuda.core.