[cudax->libcudacxx] Move device_transform to libcu++

[pciolkosz]

This PR moves device_transform from cudax to libcu++. I had to adjust some things it used from cudax. One of them was cuda::experimental::__variant, which I replaced with cuda::std::optional. I believe it was just a simpler type we can later go back to.

With device_transform moved to libcu++ there was no longer a need to keep the experimental version of copy_bytes and fill_bytes, I removed them and added device transform to the non-experimental version.

I had to still leave experimental tests of the algorithm header, because of uninitialized_buffer used in them. We can move them when we move that type.

[miscco]

I do not understand why this needed, the destructor will still run at the same time, its just that there is nothing to destroy there

[davebayer]

Design question: is device transform a good name for this object? Whenever I see the name, I am immediately thinking about some cub algorithm

[bernhardmgruber]

Same here.

[pciolkosz]

It started as launch_transform because it was used for cuda::launch API. Later we added it for copy_bytes and I believe some things Eric is working on. In general the idea was anytime execution is transferred to the device, arguments passed are going to this transformation. This is how device_transform name came up. Would device_argument_transform help with this concern? I can't think of a fundamentally different name for this

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 2h 46m: Pass: 100%/122 | Total: 1d 18h | Max: 2h 01m | Hits: 82%/236714

See results here.