Add a way to transform a CudaSlice into/from a CUdeviceptr

[Corey Lowman]

This would be an unsafe method to gain access to the underlying driver CUdeviceptr.

This would enable things like caching allocations. Notably, this would need to extract the underlying CUdeviceptr without actually dropping the CudaSlice. This can be done with std::mem::forget, but care needs to be taken to handle the host Vec if the CudaSlice owns one.