Make CuRef(Value) behave more like Ref

[maleadt]

There's some disparities between how Base.Ref(Value) behaves, and what CUDA.CuRef(Value) does:

• CuRef is separate from CuRefValue, while in Base they are linked (RefValue <: Ref)
• Ref(...) creates a constructor you can index, while calling CuRef isn't currently meant to be used outside of ccall signatures
• a CuRef isn't actually mutable, and is currently only used as a GPU-compatible counterpart of Ref for broadcasting purposes

All this currently isn't too much of a problem as we don't actually export CuRef, but it'd be nice to make it behave like Base.Ref. One possible solution would be to use unified memory, although that would disallow any conversions during CuRefValue construction (we currently call adapt).

[kshyatt]

Did #2645 partially address this?

[maleadt]

Yes, thanks for noticing.