Use a named handle type for CUDA arrays

[eyalroz]

Our CUDA array class wrapper doesn't define a handle type for arrays, rather using cudaArray* directly. Let's switch to a handle type, for better readability (it's not as though we're working with these pointers as such), better compatibility with the approach in other wrapper classes (even if they have "IDs" as handles), and better compatibility with the driver-wrappers branch, where the handle type is different.